Codex Baglama rejected every Responses WebSocket with 426, so Codex's own subscription models fell back to HTTP and re-sent the whole conversation on every step. The splitter now tunnels the WebSocket of Codex's own models to chatgpt.com byte for byte (model known from x-codex-routing-hint). Added models, remembered transfers and transfer history in a message still get 426 / a close frame and stay on the existing HTTP path, so added-model and transition handling is unchanged. bridge.json "abonelik_websocket": false restores the old refusal. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XeNgWNzPjzUKQuS32RUYXx