Claude Code /goal: why it burns the weekly quota faster

Yes. /goal sets a completion condition; after every turn a small evaluator model (Haiku by default) returns yes/no on whether the condition is met, and 'no' fires another main-model turn automatically. There is no per-turn human checkpoint, so the main model can run dozens of turns before you look up. Anthropic doubled the 5-hour cap in May 2026 (announced May 7, 2026) but left the weekly cap unchanged, so the weekly bucket is now the binding constraint for /goal sessions on Claude Code. The evaluator tokens themselves are 'typically negligible' per the docs at code.claude.com/docs/en/goal; the burn is the main model running unattended.

seven_day_oauth_apps, almost always. Claude Code authenticates via OAuth, so its requests count against the OAuth-apps weekly bucket on top of the all-up seven_day aggregate. Your account can sit at seven_day = 58% and seven_day_oauth_apps = 93% at the same time; the web UI on claude.ai/settings/usage emphasizes the all-up bar, the CLI walls on the OAuth one. The field is declared on line 24 of src/models.rs in the open-source ClaudeMeter repo.

Per the official docs, /goal is a session-scoped wrapper around a prompt-based Stop hook. The condition you typed is passed to a small fast evaluator model (defaults to Haiku) along with the conversation; the evaluator returns yes/no plus a short reason. 'No' starts another main-model turn with the evaluator's reason as guidance. 'Yes' clears the goal and returns control. The cost shape: one full main-model turn per evaluator-no, plus one tiny evaluator call per turn. With Opus as the main model, every wrong evaluator-no costs you full Opus pricing against seven_day_opus and seven_day_oauth_apps.

It fixes the 5-hour wall, not the weekly wall. The May 7, 2026 announcement doubled the rolling 5-hour Sonnet and Opus budgets for Pro and Max accounts. The weekly cap was left at the level Anthropic tightened to in 2026; it was not doubled. /goal sessions that previously hit the 5-hour wall and forced a 5-hour cooldown now keep running, which means more main-model turns charge against the same unchanged weekly bucket. Net effect: /goal users now hit the weekly wall earlier in the week than they did before the doubling, not later.

Run an ambient meter outside Claude Code. ClaudeMeter (github.com/m13v/claude-meter, MIT) polls the same /api/organizations/{uuid}/usage endpoint claude.ai/settings/usage hits, every 60 seconds (extension/background.js line 3, const POLL_MINUTES = 1), and renders the six weekly buckets plus the 5-hour bucket in the macOS menu bar. The /goal loop never sees the call. When seven_day_oauth_apps crosses 80% the popover badge changes and you can /goal clear before the next Opus turn fires. ccusage cannot do this; it sums tokens from local Claude Code JSONL files but cannot see Anthropic's server-enforced weekly utilization.

Bound the condition itself. /goal accepts up to 4,000 characters; spend some of them on an OR clause: 'The test suite passes and the diff is committed, OR you have attempted this 12 times.' The evaluator counts attempts and returns yes once the cap is hit. This costs nothing extra and stops the worst failure mode (a bad path that the evaluator keeps saying no to forever). Pair it with --model sonnet on the main model unless the goal genuinely needs Opus reasoning.

Yes, a small one. Even on a goal the model finishes in one turn, the evaluator still runs once on the resulting conversation. That call is small (Haiku-tier) and typically negligible per the docs. The cost shape only becomes significant when the evaluator returns no repeatedly and the main model fires repeatedly. Watch seven_day_oauth_apps; if it ticks up by more than 1-2 percentage points per minute during a goal, the loop is spiralling and the guard you set in the condition (or your manual /goal clear) is what saves the rest of the week.

The goal survives. Per the docs, /goal is session-scoped but persists across --resume and --continue, though the turn and token counters reset to zero on resume. The completion condition is the same; the next time you resume the session the evaluator immediately checks against it and either fires another main-model turn or clears the goal. This means a goal you forgot about can fire fresh main-model turns the next time you resume the session and resume burning seven_day_oauth_apps. /goal clear before suspending a session you do not plan to come back to.

Not by name. Anthropic's /api/organizations/{uuid}/usage endpoint reports utilization per bucket, not per session or per slash command. ClaudeMeter shows you the bucket utilization regardless of how the load got there. In practice, a /goal session shows up as seven_day_oauth_apps climbing faster than your reading speed; the rate of change is the signal. If you watch the popover for 60 seconds and the OAuth bucket jumped 2-3 percentage points, that is a /goal-shaped burn.