OmO stops failed child tasks from borrowing the parent model

OmO’s November 2 release changes what happens when a delegated coding task runs out of usable models. According to the 5.1.58 notes and merged implementation, a child task now stays on its designated category or agent model chain. It no longer borrows the parent session’s model when that chain fails.

For someone assigning different models to different jobs, this makes the assignment a more predictable boundary. A task configured for one chain should not quietly run on the model handling the main conversation after a usage limit, rejected sign-in, content-policy refusal or overload. Fallback between models already designated for the child remains possible.

When recognized refusals exhaust the chain, the implementation returns failure_kind: model_chain_exhausted. The error lists the designated models tried, their failures and any reset time supplied by the provider. An old provider_lock configuration key is now ignored because the parent fallback itself has been removed. The maintainer reports automated and live-runtime QA; AITrending has not independently executed those tests.

The scope matters: the merged pull request explicitly leaves the memory reflection worker’s separate routing behavior for another decision. This is a delegated-task change, not a promise that every background component uses the same model boundary. The release also describes browser-process cleanup fixes on macOS and Linux.

OmO is a standalone agent on senpi, a Pi fork, distinct from its older OpenCode edition. Its documented package is omo-ai; bun add -g omo-ai installs it, and the published 5.1.58 npm metadata requires Node 24 or newer. Provider credentials, subscriptions and inference costs remain separate. The source uses the Sustainable Use License, with internal-business and non-commercial restrictions, rather than an unrestricted permissive license.