Quotas, caps & failover

Subscriptions have walls: a 5-hour window, a weekly window, sometimes a per-model pool. Synergoi reads the real figures from each CLI, shows them where you're looking, lets you set caps below the vendor's, and can hand the thread to another provider the moment a wall hits.

Meters

The right rail shows a meter per connected provider; hover for windows and reset times. Each figure carries a confidence badge: measured when it came from the vendor's own report, probe when Synergoi refreshed it without spending a turn, estimated when it was derived from usage the CLI streamed. Settings › Usage & limits › Unverifiable meter data decides whether estimates are shown at all.

Caps and the watchdog

For every provider window you can set a cap below 100%, choose strict (the broker refuses to start a turn past the cap) or soft (warn, but allow), and arm a watchdog that stops a running turn when the cap is crossed mid-flight. Caps are per account, so a shared work plan can be protected while a personal one runs free.

Failover

Settings › Usage & limits › Failover has three modes:

• Off — the turn stops at the wall with a card that says why.
• Offer — the card offers the next provider in your order, with its remaining quota, and buttons to Continue, Roll back, Wait for reset or Stay.
• Auto — the hand-off happens without asking; the card records it.

A failover picks up from the turn checkpoint with a Context Pack: the original ask, progress so far, commands that already ran before the wall, the working diff and the thread's rules. The incoming model is told what it inherited and which shell it is standing in. You set the fallback order with the up/down controls in the same pane.

The usage ledger

Every turn records tokens in and out per provider and model. The footer of a thread shows the running total; the usage view in Settings charts it over time and estimates a dollar figure at public API prices (from models.dev) — not what you paid, since your subscription covers it, but a useful sense of what each thread would have cost. Cursor turns are not estimated.