DeepSeek V4: reasoning_content 400 error persists on v1.14.31 — complete fix exists in unmerged PRs

[cameronmpalmer]

Bug Description

When using DeepSeek V4 Pro or Flash models with thinking mode enabled (default), multi-turn conversations fail with:

The `reasoning_content` in the thinking mode must be passed back to the API.

This happens reliably on v1.14.31 (latest release) in conversations with multiple tool-call turns, regardless of whether the conversation is fresh or resumed from history.

Why the existing fixes are insufficient

The release notes mention:

• v1.14.24: "Fixed DeepSeek assistant messages so reasoning is always included"
• v1.14.29: "DeepSeek OpenAI-compatible setups now keep reasoning_content interleaved by default"

These handle some code paths but not all. The fundamental issue is that reasoning_content must be present on every assistant message in the conversation history — historical messages from DB replay, string-content messages, and messages processed on the second interleaved pass all need it. The current code has gaps.

The fix exists — three separate PRs wrote it

Three contributors have submitted PRs with the complete fix:

1. fix(provider): complete DeepSeek reasoning_content round-trip for multi-turn conversations #24250 by @knefenk — Covers all three layers: auto-enable interleaved, dynamic SDK key, and unconditional reasoning_content injection for all assistant messages including DB-replayed ones. Commit: 41eb35a
2. fix: preserve existing reasoning_content on second interleaved pass (#24146 follow-up) #24428 by @claudianus — Fixes second-pass regression where empty reasoningText overwrites preserved providerOptions values. Commit: 86dd22f
3. fix: generalize reasoning_content injection across AI SDK providers #24895 by @Malkovich-666 — Generalizes the fix across all AI SDK providers with a providerOptionsKey() helper. Commit: c42e105

All three were closed without merge.

Environment

• OpenCode v1.14.31 (Linux binary)
• Provider: @ai-sdk/openai-compatible via LLM Gateway (pinned to DeepSeek API)
• Model: deepseek-v4-pro with reasoning: true and interleaved: { field: "reasoning_content" }
• Also reproduced with direct DeepSeek API (bypassing gateway)

Request

Please review and merge one of the existing complete fixes (#24250, #24895) so that DeepSeek V4 with thinking mode works reliably in multi-turn conversations.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors #24722: DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors (open, very similar)
• deepseek V4 flash 或者pro 出现 链接中断 ，中断后 再也不能正常对话工作了Error from provider (DeepSeek): The reasoning_content in the thinking mode must be passed back to the API. #24920: DeepSeek V4 flash/pro reasoning_content error after conversation interruption (open)
• deepseek v4 pro开启前思考模式的时候api层丢弃reasoning_content导致接口调用失败 #24714: DeepSeek v4 pro thinking mode dropping reasoning_content causing API failure (open)
• reasoning_content stripped from assistant messages on replay, causing KV cache invalidation on local inference #19081: reasoning_content stripped from assistant messages on replay (open)
• Thinking blocks in assistant messages are modified/dropped, causing API error on multi-turn conversations #18254: Thinking blocks in assistant messages modified/dropped, causing API error on multi-turn conversations (open)

There are also many closed issues tracking this same error (e.g. #24261, #24442, #24190, #24803). The maintainers are likely already tracking this bug — please check the above issues and consider adding a +1 or relevant details there rather than opening a new report.

[chrispository]

I can confirm the first one you linked fixes it: #24250

I had OpenCode pull the main dev branch and asked it apply the linked fix. Then I re-built, and it's working perfectly for me now. You'll just have to build it, took less than 20 seconds for me.

https://github.com/chrispository/opencode_deepseek_fix