Background AI (engine, who pays, cost per job)
One Settings panel listing every automatic job Omniscio runs on its own — master-debt fixers, Help Desk investigations, auto-lander rescues, cloud dispatcher workers, daily maintenance and the AI helpers — with the engine each one uses, who pays for it, whether it counts against your Claude plan limits, and what a run really costs, all switchable in one click or by preset.
Omniscio runs several jobs without being asked. Each already had its own engine setting, scattered across different screens, and none of them said who was paying. Settings → Background AI puts them in one list.
What each row tells you
- Engine — the model the job runs on, changeable in one click. This writes the same setting the job already reads, so the panel and the job can never disagree.
- Who pays — the company lane, your Claude plan, your own API key, or Omniscio credits. This is the app's one payer rule, applied to that job with the engine it is on: a DeepSeek job can never be reported as running on your Claude plan, because that lane cannot serve a DeepSeek model.
- Plan limits — whether running the job counts against your Claude subscription's 5-hour and weekly limits, or not.
- Cost per run — the average of that job's own last 20 runs, read from the spend ledger. A job that has never run says "No runs recorded yet" rather than showing a number.
Presets
- Free to me — the company lane's model when your account is company-funded, otherwise the cheap Claude tier. It never quietly charges credits.
- Cheapest — DeepSeek V4.1 Flash, and the DeepSeek key first for the helpers.
- Best quality — your own default Claude model.
A preset changes the same settings the rows change. Nothing else is stored.
A lane outage does not stop a job
When a job's chosen engine cannot serve for a moment, the job runs on a free option you already have instead of stopping — never the wallet, and never your credits on a company-funded account. When the refusal is a model your company lane does not run, the card names the job that was refused and points at Background AI settings, so you can tell what stopped instead of only which model was named.
Which jobs are in the panel
Master-debt fixers, Help Desk investigations, auto-lander rescues, cloud dispatcher workers, and the AI helpers (session titles, suggested replies, reply drafts, summaries).
Three automatic jobs are deliberately absent for now: daily maintenance (its engine needs a provider and a model written together, which this panel's one-click switch does not do — use Dev Pipeline → Setup), the master-sync escalator, and the worktree AI triage — the last two because neither marks the sessions it starts, so the ledger cannot tell their runs apart and a cost figure would be a guess.