[Bug] DeepSeek V4 thinking-mode round-trip fails deterministically — 100% reproducible on turn 2 and after context compaction

[yingpinggan]

When using deepseek-v4-pro with thinking/reasoning mode (the default), Moon Bridge fails to correctly preserve the assistant's thinking content block when constructing the next outgoing request. DeepSeek rejects the malformed request with a 400 error, and the client (Codex) then retries the identical broken request indefinitely — the conversation never recovers without starting a new session.

This is not intermittent. It is deterministic and reproducible at two distinct trigger points.

Error

[WuzeiMan]

The content[].thinking in the thinking mode must be passed back to the API.

同问，在思考模式下，content[]中的thinking字段必须被传回API