DeepSeek V4 Pro (OpenRouter) reasoning_content error - v1.14.27 #24569

Activity

1. 077dwei

   added

   bugSomething isn't workingSomething isn't working

   on Apr 27, 2026
2. github-actions

   assigned

   rekram1-node

   on Apr 27, 2026
3. github-actions

   added

   coreAnything pertaining to core functionality of the application (opencode server stuff)Anything pertaining to core functionality of the application (opencode server stuff)

   on Apr 27, 2026
4. github-actions commented on Apr 27, 2026

   

   github-actionsBot

   on Apr 27, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of several existing issues reporting the same reasoning_content error with DeepSeek thinking mode. Please check:

   • The bug requiring deepseek requests to include the thinking content still exists in versions 1.14.26 and 27. #24558: Same bug still present in v1.14.26 and v1.14.27
   • deepseek v4 provider error: The reasoning_content in the thinking mode must be passed back to the API #24497: DeepSeek V4 reasoning_content provider error (same error message)
   • DeepSeek V4: reasoning_content lost on second interleaved pass (regression from #24146) #24442: DeepSeek V4 reasoning_content lost on second interleaved pass (regression)
   • DeepSeek V4 400: The reasoning_content in the thinking mode must be passed back to the API. #24385: DeepSeek V4 400 error — same error message
   • Bug: DeepSeek reasoning_content not passed back in multi-turn conversations #24261: DeepSeek reasoning_content not passed back in multi-turn conversations
   • [DeepSeek] The reasoning_content in the thinking mode must be passed back to the API. #24188: [DeepSeek] The reasoning_content in the thinking mode must be passed back to the API
   • DeepSeek V4 Flash: reasoning_content must be passed back to API when using thinking mode #24130: DeepSeek V4 Flash — same reasoning_content error
   • The reasoning_content in the thinking mode must be passed back to the API. #24083: The reasoning_content in the thinking mode must be passed back to the API

   If your issue is the same, please consider adding a reaction or comment to one of the above instead of opening a new report.
5. rekram1-node commented on Apr 27, 2026

   

   rekram1-node

   on Apr 27, 2026

   Collaborator

   More actions

   @077dwei please please provide a session demonstrating it:

   opencode export > session.json

   RUN THIS IN YOUR TERMINAL DO NOT USE /export

   And then send me the session dump.

   👍React with 👍1Reacted by d770
6. Meepoljdx commented on Apr 27, 2026

   

   Meepoljdx

   on Apr 27, 2026

   More actions

   same issue @rekram1-node this is my session json

   session.json

   👍React with 👍1Reacted by d770
7. 077dwei commented on Apr 27, 2026

   

   077dwei

   on Apr 27, 2026

   Author

   More actions

   @rekram1-node

   session.json

   Trimmed session (only error-related part):

   👍React with 👍1Reacted by d770
8. rekram1-node commented on Apr 27, 2026

   

   rekram1-node

   on Apr 27, 2026

   Collaborator

   More actions

   @Meepoljdx you are using something called:
   "providerID": "deepseek-plat",

   I need to see your provider config it's prolly missing interleaved settings

   👍React with 👍1Reacted by d770
9. broaz27 commented on Apr 27, 2026

   

   broaz27

   on Apr 27, 2026

   More actions

   I can't reproduce the issue anymore, but only in new sessions, im on deepseekv4 pro, provided by deepseek on openrouter, my interleaved settings are on their default "reasoning_content"
10. rekram1-node commented on Apr 27, 2026

    

    rekram1-node

    on Apr 27, 2026

    Collaborator

    More actions

    @077dwei I can't have a trimmed session, I need your actual session otherwise it provides basically 0 value. If you need it secret use --sanitize

    The session you sent has 0 assistant messages so the error would be impossible
11. 077dwei commented on Apr 27, 2026

    

    077dwei

    on Apr 27, 2026

    Author

    More actions

    @rekram1-node Sorry for sending the trimmed version earlier. Here is the full session exported with --sanitize:

    session_sanitized.json

    👍React with 👍1Reacted by d770
12. ngkhanhlinh2k4 commented on Apr 27, 2026

    

    ngkhanhlinh2k4

    on Apr 27, 2026

    More actions

    me too, still error like that
13. rekram1-node commented on Apr 27, 2026

    

    rekram1-node

    on Apr 27, 2026

    Collaborator

    More actions

    @ngkhanhlinh2k4 more info plz
14. yujf commented on Apr 27, 2026

    

    yujf

    on Apr 27, 2026

    More actions

    session.json
    If a shell command is executed even once in the middle, a 400 error will occur.
    @rekram1-node
15. 372572571 commented on Apr 27, 2026

    

    372572571

    on Apr 27, 2026

    More actions

    Actually, as long as you use the OMO plugin's Plan Builder with the DeepSeek V4 model to start planning a feature, when your context usage reaches about 10%, this problem will inevitably be triggered — it feels like something has been cleared.

    👍React with 👍1Reacted by kenlive
16. 3 remaining items

    Load more

17. rekram1-node commented on Apr 27, 2026

    

    rekram1-node

    on Apr 27, 2026

    Collaborator

    More actions

    For anyone using custom providers do this:

    "provider": {
        "custom-deepseek": {
          "models": {
            "deepseek-v4-pro": {
              "interleaved": {
                "field": "reasoning_content",
              },
            },
          },
        },
      },
    

    I'll configure stuff to default it to set this for deepseek models but for now you need to follow this as mentioned in many issues and many prs (@yujf @Meepoljdx)

    👍React with 👍4Reacted by d770, Kay, yujf and EwardWang

    🎉React with 🎉1Reacted by Kay
18. yujf commented on Apr 28, 2026

    

    yujf

    on Apr 28, 2026

    More actions

    Thank you, it works fine after adding interleaved. @rekram1-node
19. hearkenmx commented on Apr 28, 2026

    

    hearkenmx

    on Apr 28, 2026

    More actions

    Same issue here. Environment details:

    • opencode v1.14.28
    • Provider: @ai-sdk/openai-compatible with custom endpoint (myeth)
    • Model: deepseek-v4-pro
    • Error: AI_APICallError: The reasoning_content in the thinking mode must be passed back to the API.

    Occurs consistently on the 2nd+ turn. Workaround: setting "reasoning": false in opencode.json model config.
20. matteo-pacini

    added a commit that references this issue on Apr 28, 2026

    feat(opencode): research-backed agent tuning (#225)

    ...

    fa42662
21. rekram1-node commented on Apr 28, 2026

    

    rekram1-node

    on Apr 28, 2026

    Collaborator

    More actions

    @hearkenmx refer to what I said above, or provide your config

    👍React with 👍1Reacted by Max
22. jukefr commented on Apr 28, 2026

    

    jukefr

    on Apr 28, 2026

    Last edited by jukefr

    More actions

    still having issues with this even after doing what you suggested with interleaved....

    {
      "$schema": "https://opencode.ai/config.json",
      "permission": {
        "*": "allow",
        "doom_loop": "ask",
        "external_directory": "ask",
        "read": {
          "*": "allow",
          "*.env": "deny",
          "*.env.*": "deny",
          "*.env.example": "allow"
        },
        "bash": {
          "*": "allow",
          "rm -rf *": "ask",
          "git push --force*": "deny",
          "git push -f *": "deny",
          "DROP TABLE*": "deny",
          "DROP DATABASE*": "deny"
        }
      },
      "mcp": {
        "context7": {
          "type": "remote",
          "url": "https://mcp.context7.com/mcp",
          "enabled": true
        },
        "engram": {
          "type": "local",
          "command": [
            "engram",
            "mcp",
            "--tools=agent"
          ]
        }
      },
      "snapshot": false,
      "autoupdate": "notify",
      "share": "manual",
      "provider": {
        "openrouter": {
          "models": {
            "deepseek/deepseek-v4-flash": {
              "interleaved": {
                "field": "reasoning_content"
              }
            }
          }
        }
      }
    }
    

    the first like 5 requests seem to go fine but then the

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

    appears out of nowhere

    it happened right after a todowrite tool call if that helps ?

    and also the session appear to be unrecoverable from this point onwards no matter what prompt you pass to deepseek

    and im running latest release also 1.14.28
23. i-blis commented on Apr 28, 2026

    

    i-blis

    on Apr 28, 2026

    Last edited by i-blis

    More actions

    Issue persists even with

    "provider": {
        "openrouter": {
          "models": {
            "deepseek-v4-flash": {
              "interleaved": {
                "field": "reasoning_content",
              },
            },
          },
        },
      },
    

    Running 1.14.28, session posted above.
24. rekram1-node commented on Apr 28, 2026

    

    rekram1-node

    on Apr 28, 2026

    Collaborator

    More actions

    Yeah it's a bug in openrouter sdk ill get to it
25. DiegoMagdaleno commented on Apr 28, 2026

    

    DiegoMagdaleno

    on Apr 28, 2026

    More actions

    Having the same issue with OpenCode Go, I reproduced it on a random task.

    session.json
26. rekram1-node commented on Apr 28, 2026

    

    rekram1-node

    on Apr 28, 2026

    Collaborator

    More actions

    @DiegoMagdaleno that's because you are on an older version, please upgrade
27. rekram1-node commented on Apr 28, 2026

    

    rekram1-node

    on Apr 28, 2026

    Collaborator

    More actions

    Okay OpenRouter said they are merging a fix on their end

    Note to those in this thread this is an issue about OPENROUTER so if you are having issues with a specific provider check:

    • are you on latest
    • are you using openrouter? If not open a new issue or comment on separate one
    • are you using custom provider? If so be sure to set the interleaved reasoning fields

    👍React with 👍1Reacted by XaPoHbomj
28. rekram1-node commented on Apr 28, 2026

    

    rekram1-node

    on Apr 28, 2026

    Collaborator

    More actions

    Okay their change rolled out to their api, I tested and was unable to reproduce issue after, will close this but continue monitoring

    🎉React with 🎉1Reacted by Kay
29. rekram1-node

    closed this as completedon Apr 28, 2026
30. hearkenmx commented on Apr 29, 2026

    

    hearkenmx

    on Apr 29, 2026

    More actions

    new-api based router has the same issue, in case anyone use that infra, I think they already aware of it.
    QuantumNous/new-api#4408

    👍React with 👍1Reacted by dream zhao
31. 0x7CFE commented on May 27, 2026

    

    0x7CFE

    on May 27, 2026

    Last edited by 0x7CFE

    More actions

    Experienced the same issue today on OpenRouter. Everything just broke. OTOH, free tier model from Zen worked flawlessly.

    Tried switching/banning specific providers using OpenRouter guardrails, got this or just 400.

    

    P.S.: Also usage price just exploded. Don't sure if it's related. Looked almost like queries weren't cached at all.

    👍React with 👍4Reacted by Peter Mills, Mikhail, Nifury and Quentin Gillet