First Oh My Pi run made 1,051 model calls, drained 3 subscriptions, then stalled

A first-time user of Oh My Pi (OMP), a multi-agent orchestration tool, assigned it one large repository refactoring task. The intended setup split roles across models: GPT-5.6 Sol Max for planning, Grok 4.5 for coordinating implementation, Composer 2.5 as the worker doing the actual coding, and GPT-5.6 Sol Medium for limited review passes.

Instead, the run lasted 5 hours 42 minutes, triggered 1,051 separate model calls, and drew down usage from three different subscription pools simultaneously, while leaving the task mostly unfinished. It ended with the error "Provider stream stalled while waiting for the next event" — the user did not manually stop it; it froze on its own while a Composer worker was still active.

In that single run, roughly 54% of the weekly Codex quota, 24% of the weekly Grok/xAI quota, and 13% of the monthly Cursor usage were consumed (these are separate products with different quota windows, so the percentages shouldn't be added together). By comparison, the same user's usual direct Codex-only workflow — handling both planning and implementation — typically consumed only about 8% of the weekly Codex quota.