The repository description states that MiniMax-AI/OpenAgentCore is an open-source self-hosted Agents API implementation that runs Codex, Claude Code, and MiniMax Code with the official OpenAI SDK on its own models and its own machines. The November 5, 2026 (UTC) commit ff41380(#920) expanded the settings that Session updates in this implementation can receive. A Session is the unit of work in which an agent drives a model, and a single request handling within it is called a Turn; until now, the contract document stated that updates received only metadata. Now they receive agent.model, agent.reasoning, agent.service_tier, and review, and reject agent.tools.
What is actually applied is determined by each Harness. A Harness is an adapter that attaches to an execution entity such as Codex or Claude Code, and this commit added Reviews and SessionUpdates slots to the Harness declaration contract. SessionUpdates enumerates Model, ReasoningEffort, ReasoningEffortReset (the case of returning effort to null), and ServiceTier with their respective support flags, and Core rejects updates that change unsupported settings, while supported changes are passed to the adapter as the configuration for the next Turn. For review, only auto_review may remain in the list once none is excluded. Updates that change only metadata no longer re-validate the Session configuration, and the backfill that populates the settings columns was explicitly stated to be unnecessary because there is no in-place upgrade. The runtime settings query response was changed to display the model changed within the session as source session.
However, in the Harness capability table, the cases of changing the model, reasoning effort (including null), service tier, and review: auto_review via a Session update, as well as the case of changing agent.tools, are all recorded as Rejected in every column. What the columns of the table mean was not confirmed in this change. The contract document, the Harness declaration files (claudesdk, codex, mcode), and the core API handling code were changed together, and the changed line count is 867 lines (+647/-220). No performance or stability metrics were presented. Since the use case for why switching models mid-session is needed was not included in the commit, it seems realistic to check each Harness's declarations and the capability table first before using it.