v2.43.3

Compatible Reasoning Across Model Switches

Preserve useful reasoning context when switching models without exposing incompatible reasoning as ordinary assistant output.

Fixed

• Compatible structured replay. Recognized OpenAI-compatible reasoning fields remain structured when switching between compatible reasoning models on the same provider.
• Clearly labeled fallback. Readable reasoning from incompatible models is wrapped in <reformatted-pre-switch-reasoning> markers instead of appearing as an ordinary prior answer.
• Opaque state protection. Redacted and encrypted-only reasoning is omitted from incompatible requests.
• Safe return switching. Outbound conversion does not mutate session history, allowing the original model to reuse its protocol state when switching back.
• Non-reasoning destinations. Models without reasoning support receive the labeled plaintext fallback instead of unsupported structured fields.

Technical

• Restricts structured reasoning properties to reasoning_content, reasoning, and reasoning_text.
• Adds regression coverage for Kimi model switching, cross-provider and cross-API handoffs, parser behavior, separator formatting, and history preservation.
• Updates model-switching documentation across the root, coding-agent, and AI package guides.

Reproduction and OAuth testing contributed by @Acters in issue 374. Implemented in PR 386.