Fixes multi-step note tools failing with ChatGPT OAuth and improves provider history compatibility.
- Correct assistant text replay and preserve tool-call/result pairs through Stop, Clear, reload and long conversations.
- Preserve Anthropic thinking signatures, citations and server search; handle pause_turn and adaptive thinking for new Claude models.
- Isolate OpenAI response chaining and native model/account history.
- Fetch authenticated model catalogs for all providers with a persistent 24-hour cache, background refresh on settings visits and immediate manual refresh. Failed requests keep the previous list and selection.
- Discover the latest stable Codex client version and use available reasoning/capability metadata. Preserve custom model IDs, including GPT-6.1, after reload.
- Prevent late request failures from affecting a new conversation, freeze provider settings per turn, and disable default transcript debug logging.
Validated with production-module regression tests, TypeScript/Svelte checks and production build. Live OAuth testing successfully called eight listed models, including GPT-6.1 Sol, and completed GPT-5.5/GPT-6.1 multi-step synthetic tools. Live testing used an existing Codex OAuth session; physical iPhone, plugin Device Flow and live Anthropic testing were not exercised.