Limits & Notes

Item	Limit
Max request body	100 MB (including multimodal base64 images)
Streaming read timeout	300s (single response disconnects after 5 minutes)
WebSocket (voice S2S)	24h long-connection timeout
Concurrent requests	Codex / Cursor plans run one task per connection and enforce global / per-admin queue limits. Apply QPS control to DeepAgents too, to reduce SQLite write-lock contention.
Single-session history length	Bounded by underlying model context window; very long sessions should be split into new conversations
`conversation_id` cross-instance	Not portable — valid only within this deployment instance
Key-level isolation	One `sk-svc-` key maps to one service; conversations/data are fully isolated across services