Deepseek v4 thinking/reasoning text

[druellan]

For some reason, often the Deepseek v4 models tend to display the reasoning/thinking text directly in the chat window:

Looks like the reasoning/thinking tags are being filtered, but I played with the extension options, tesed new projects and new profiles and still can't understand why this happens.

Anyone is having this issue?
Thanks!

[divyeshio]

@ltmoerdani Please release the new version as soon as possible!

[weizhen25]

same problem. the answer is too long with too much thinking/reasoning text. please hide it. or a option to handle it.

[druellan]

Yeah, for me is also still broken. Sometimes it work (the thinking happening inside the VSCode thinking block), the next session the thinking is just dropped into the chat as is, and sometimes even tool calls are dropped inside the chat.

[weizhen25]

do you have a fix or workaround to avoid that? the thinking text is too long for pageup and summary review.

[druellan]

I do not, it is very random, so I'm still not sure why sometimes fail. For example right now it is working fine. I have this settings enabled:

"extensions.supportAgentsWindow": {
    "ltmoerdani.opencode-copilot-chat": true,
  },
  "opencodego.inlineSuggestionsModel": "deepseek-v4-flash",
  "opencodego.autoEnableAgentsWindow": true,

That I don't think are realted, but I'll leave them here just in case.

[druellan]

I don't see this happening on the new version, so, seems solved. I'll reopen if I see the issue again.

[ltmoerdani]

Thanks for the detailed discussion @druellan, glad the latest version seems to have sorted it out on your end.

i went through this after you posted, tracing through git history and all the related thinking/reasoning fixes. turns out this was an accumulation of a few things that all produced similar symptoms:

1. gateway bug: the Go gateway wraps all output in reasoning_content instead of content when thinking is off (upstream #37635). the workaround depends on chunk ordering, which explains the intermittent nature.
2. multi-turn echo missing: before PR fix(history): echo reasoning_content on multi-turn thinking requests #123, DeepSeek's reasoning_content wasn't echoed on follow-up turns, causing HTTP 400 errors that could trigger fallback behavior.
3. regression from refactor: the god file split in PR refactor: split god files into domain modules + data-driven model registry #155 changed how flushReasoningFallback() works, which could cause reasoning to either get silently dropped or leak as visible text depending on context.

i documented the full analysis in #196 if anyone wants the gory details. the short version: all three are addressed in current builds (0.7.x), so it should stay stable going forward.

feel free to reopen if it comes back. would be good to know what model/thinking-effort combo triggers it so we can narrow down whether it's the gateway workaround or something else.