OpenAI Python SDK 3.42.0 adds agent context compaction

SDK 3.42.0 moves agent controls from server-side defaults into client code. The release notes frame this explicitly: summarization behavior that used to be opaque is now something you set in your own call, alongside agent session review and tool settings and experimental agent time awareness. Those three agent-focused additions sit next to security work in the same version, including organization external key management and callable API keys scoped to each request — a pairing that suggests credential handling and agent behavior are being updated together rather than as separate tracks.

The rest of the release touches real-time and WebSocket plumbing: SIP DTMF commands and delivery events, improved WebSocket recovery and delivery reporting, WebSocket warmup with preserved cleanup cancellation, and a fix that preserves refusal output and authentication in WebSocket helpers. For agent builders, the compaction setting is the first item to read, since it changes what the client can control directly.