$ the-wire · showcase
openai-node and openai-python ship managed Responses WebSocket sessions
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
![openai-sdks[bot]](https://avatars.githubusercontent.com/in/3705508?v=4)
Both OpenAI SDKs picked up managed Responses WebSocket sessions, while codex tightened sandbox and network policy validation and the deprecation clock started on MCP connector_id.
feat(api): add managed Responses WebSocket sessions openai/openai-node
openai-node now runs Responses over managed WebSocket sessions: client headers and per-connection overrides survive the handshake, independent response lanes are routed and final responses collected, and reconnects are explicit with application-level restoration and no automatic request replay. The same surface lands in openai-python.
feat(api): add managed Responses WebSocket sessions openai/openai-python
The Python SDK gains the same managed Responses WebSocket support, validated against local WebSocket regression suites covering custom request headers, independent lanes, and connection lifecycle.
chore(api): deprecate MCP connector_id (#3894) openai/openai-python
MCP tool docs in Responses, beta Responses, and Realtime now carry deprecation guidance for connector_id on models released after September 1, 2026, pointing at server_url for remote MCP servers and tunnel_id for Secure MCP Tunnels. The connector field and its request/response behavior stay available, so existing code keeps working while the migration path is documented.
Validate network socket policies using the executor OS openai/codex
Network policy validation now runs against the executor's OS rather than the controller's, because the two can differ and absolute paths valid on the executor (Windows paths under a Linux controller) were being rejected. Allowed socket paths must be NUL-free and absolute for the executor OS.
Repair Windows sandbox access to existing runtime children openai/codex
Windows sandbox access is repaired for runtime children that were never granted inherited permissions: Codex walks the app runtime tree, adds missing read/execute grants for the sandbox group with GRANT_ACCESS to preserve existing grants, and validates access afterward.
Preserve bio policy errors as a distinct non-retryable error openai/codex
Streaming bio_policy failures were being classified as generic invalid requests; Codex now carries a distinct BioPolicy error across the API and core protocol, recognizes HTTP 400 responses and wrapped WebSocket errors, preserves server messages, and treats these as non-retryable. Also out: reasoning effort is preserved for synchronous Guardian reviews, and the two Python and Node SDK releases ...
Action items
- → Migrate MCP tool configs off connector_id to server_url (remote servers) or tunnel_id (Secure MCP Tunnels) for models... openai/openai-python [plan]
