reasoning: v2.0.19 + opencode-go/deepseek-v4.1-flash — replies hidden inside reasoning, and CoT fragments leaked into text #54861

Activity

1. opencode-agent commented on Oct 12, 2026

   

   opencode-agentBot

   on Oct 12, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report and the database analysis.

   I couldn't test this against the real opencode-go gateway because there are no provider credentials here. Instead, I ran OpenCode 2.0.19 with a simulated model:

   • Answer sent only as reasoning, with no text: the TUI shows just a collapsed "Thought" block, like your Direction A turn. OpenCode is showing exactly what it received.
   • Reasoning that quotes reasoning: opencode zen (big-pickle) inline </think> leaks into text content instead of a reasoning part #54599's title, literal </think> included, followed by a normal answer: the whole quoted line stayed inside the Thought block, and only the answer was shown as text. OpenCode 2.0.19 doesn't split text at </think> for OpenAI-compatible providers.

   All three of your Direction B text parts start right after the </think> in #54599's title ( leaks into text content instead of a reasoning part"). Since OpenCode doesn't split there, the cut most likely happens upstream: the gateway or model server seems to treat a quoted </think> in the reasoning as the end of reasoning and send the rest as normal text. The reasoning-only turns fit the same picture: #54324 saw them on 7.3% of opencode-go turns and 0% on the official DeepSeek API, with the same model.

   This needs a maintainer with access to the OpenCode Go gateway. A raw SSE capture of a failing turn (delta.reasoning_content and delta.content) would show for certain where the answer ends up.

   Related issues, most relevant first:

   • opencode-go / deepseek-v4.1-flash: the finished answer is returned in reasoning_content while content is empty (7.3% of turns; 0% on the official API in the same session) #54324: same provider and model, answer returned in reasoning_content, gateway comparison (closed as a duplicate of [BUG] OpenCode Go DeepSeek V4.1 Flash intermittently returns reasoning-only responses after model ID migratio #48419)
   • [BUG] OpenCode Go DeepSeek V4.1 Flash intermittently returns reasoning-only responses after model ID migratio #48419: v1 report of reasoning-only turns on opencode-go/deepseek-v4.1-flash
   • DeepSeek V4.1 Flash: answer emitted only in reasoning (output=0, finish=stop) and reasoning leaks into visible text #49673: v1 report of both symptoms with DeepSeek V4.1 Flash, with SSE captures from normal turns
   • reasoning: opencode zen (big-pickle) inline </think> leaks into text content instead of a reasoning part #54599: inline </think> leaking into text with big-pickle (a different model)
2. chfabbro commented on Nov 7, 2026

   

   chfabbro

   on Nov 7, 2026

   More actions

   Same split on the OpenRouter lane, not just opencode-go.

   OpenCode 2.2.5/2.2.6, provider openrouter, model ~deepseek/deepseek-flash-latest (routed to deepseek/deepseek-v4.1-flash-1017). A prompt that makes the model write the literal thinking tag is enough. The provider ends delta.reasoning mid-token and continues the same chain-of-thought on delta.content:

   delta.reasoning: "... clients normalize thinking by wrapping in ` thinking..."
   delta.content:   "` tags. The question likely references a specific provider implementation..."
   

   Usage for that response: completion_tokens=5974, reasoning_tokens=180, so 5,794 tokens of visible output for 11.9 KB of CoT. In a longer agent session this reached ~94 KB before the output limit.

   OpenCode stores what it receives. In a proxy run, 12/12 assistant text parts were byte-identical to the joined delta.content, and the reasoning parts matched the joined delta.reasoning.

   This looks like the same mechanism as #55268 (big-pickle), except the trigger here is an opening tag inside the reasoning on DeepSeek rather than a closing tag.

   Repro request

   request.json is the 176-token synthetic prompt:

   {
     "model": "~deepseek/deepseek-flash-latest",
     "messages": [
       {
         "role": "user",
         "content": "We are debugging a chat application's Ollama provider. Think through this carefully and at length.\n\nQuestion: when Ollama returns a qwen3 response with `message.thinking`, the provider wraps that thinking in a literal ` thinking` tag. Explain exactly how that wrapping works, whether the literal string ` thinking` can end up inside `message.content`, and what a downstream consumer sees when the tags leak.\n\nRules:\n- Reason step by step before answering.\n- When you mention the opening tag, write it exactly as ` thinking` (backtick, less-than, the word thinking, greater-than, backtick) every single time.\n- Do not summarize early; explore the code paths and edge cases in detail."
       }
     ],
     "stream": true,
     "max_tokens": 6000,
     "usage": {
       "include": true
     }
   }

   Raw SSE around the boundary

   16 chunks; generation IDs redacted:

   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":"3 models","reasoning_details":[{"type":"reasoning.text","text":"3 models","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":" use","reasoning_details":[{"type":"reasoning.text","text":" use","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":" thinking. Some","reasoning_details":[{"type":"reasoning.text","text":" thinking. Some","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":" clients","reasoning_details":[{"type":"reasoning.text","text":" clients","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":" normalize","reasoning_details":[{"type":"reasoning.text","text":" normalize","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":" thinking by","reasoning_details":[{"type":"reasoning.text","text":" thinking by","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":" wrapping in ` thinking","reasoning_details":[{"type":"reasoning.text","text":" wrapping in ` thinking","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"","role":"assistant","reasoning":"...","reasoning_details":[{"type":"reasoning.text","text":"...","format":"unknown","index":0}]},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":"` tags. The","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":" question likely references","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":" a specific provider","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":" implementation.","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":" We need explain","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":" exactly how wrapping","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":" works","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: {"id":"gen-REDACTED","object":"chat.completion.chunk","created":1794087676,"model":"deepseek/deepseek-v4.1-flash-1017","provider":"Alibaba","choices":[{"index":0,"delta":{"content":",","role":"assistant"},"finish_reason":null,"native_finish_reason":null}]}
   
   data: [DONE]
   

   Trimmed usage

   {
     "id": "gen-REDACTED",
     "object": "chat.completion.chunk",
     "created": 1794087676,
     "model": "deepseek/deepseek-v4.1-flash-1017",
     "choices": [
       {
         "index": 0,
         "delta": {},
         "finish_reason": "length",
         "native_finish_reason": "length"
       }
     ],
     "usage": {
       "prompt_tokens": 176,
       "completion_tokens": 5974,
       "total_tokens": 6150,
       "completion_tokens_details": {
         "reasoning_tokens": 180
       }
     }
   }

   Repro:

   curl -sS -N https://openrouter.ai/api/v1/chat/completions \
     -H "Authorization: Bearer $OPENROUTER_API_KEY" \
     -H "Content-Type: application/json" \
     --data-binary @request.json