Bug: Assistant response prefill is incompatible with enable_thinking.: {"error":{"code":400,"message":"Assistant response prefill is incompatible with enable_thinking.","type":"invalid_request_error"}}

[akierum]

Describe the bug

I have made a task in aider-desk. for some reason after some thinking this error is shown.

Assistant response prefill is incompatible with enable_thinking.: {"error":{"code":400,"message":"Assistant response prefill is incompatible with enable_thinking.","type":"invalid_request_error"}}

Steps to reproduce

Open aiderdesk Version: 0.65.0
Use Qwen 3.6 27B Q8 community version

Expected behavior

No thinking error.

Operating System

Windows 11

Version

Version: 0.65.0

Screenshots

attached

[wladimiiir]

Thanks for the report.

What provider is this?

[akierum]

Jan.ai latest version v0.7.9.
model qwen 3.6 27B Q8 community version.

[wladimiiir]

This might be an issue of some specific extension. What extensions do you have installed? Can you try disable all of them to see if the problem persists and then identify which one is causing it? Thanks.

[akierum]

Hello, happens without extensions too.

…

On Tue, May 12, 2026 at 10:17 PM Vladimir Hrusovsky < ***@***.***> wrote: *wladimiiir* left a comment (hotovo/aider-desk#785) <#785 (comment)> This might be an issue of some specific extension. What extensions do you have installed? Can you try disable all of them to see if the problem persists and then identify which one is causing it? Thanks. — Reply to this email directly, view it on GitHub <#785 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHQVDIRBQMLXWFFMQ3ST6JT42N2DLAVCNFSM6AAAAACY263ZBGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DIMZTHEYDMOJWG4> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[wladimiiir]

I will need more information, as without it, it's not possible to tell you what the problem is.

[akierum]

Hello, re-prompting solves the issue so I cant know anything more. It happens sometimes.

…

On Wed, May 13, 2026 at 6:48 PM Vladimir Hrusovsky ***@***.***> wrote: *wladimiiir* left a comment (hotovo/aider-desk#785) <#785 (comment)> I will need more information, as without it, it's not possible to tell you what the problem is. — Reply to this email directly, view it on GitHub <#785 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHQVDIXCBZ2L3O5ZBYWOIUD42SKNPAVCNFSM6AAAAACY263ZBGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DINBSG44DAOBTG4> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[wladimiiir]

This error (Assistant response prefill is incompatible with enable_thinking) does not originate from AiderDesk — it is returned by the llama.cpp inference engine that powers your Jan.ai backend. Jan.ai uses llama.cpp (via cortex.llamacpp) under the hood for serving GGUF models.

The llama.cpp server enforces a constraint: when enable_thinking is active (which is the case for thinking-capable models like Qwen 3.5/3.6), the last message in the API request cannot have role: "assistant" (known as "assistant response prefill"). This can get triggered during multi-turn tool-use conversations where the message history naturally ends with an assistant message. This is a known upstream issue — see ggml-org/llama.cpp#20861 and mastra-ai/mastra#15234 for the same error with Qwen models.

To help me investigate further:

1. Which mode were you using? — Agent mode, or one of the Aider modes (Code, Ask, Architect, etc.)?
2. Does re-sending the same prompt fix it? If it's intermittent, it may be triggered by specific conversation patterns (e.g., after tool calls).
3. Could you share the context.json file from the project's .aider-desk/tasks/{taskId} directory? This would show me the exact message sequence being sent. (You can sanitize any sensitive content.)
4. What were the last few messages before the error appeared? (e.g., was it after a tool call, after a long conversation, right from the start, etc.)

As a potential workaround, you could try disabling thinking/reasoning mode in Jan.ai's model settings if available, or check if your Jan.ai version exposes any llama.cpp flags related to prefill-assistant.

[akierum]

*Which mode were you using* *qwen 3.6 27b q8 lmstudio community edition* *Does re-sending the same prompt fix it?* *sometimes fix it.* Could you share the *context.json* file from the project's .aider-desk/tasks/{taskId} directory? attached rar file What were the *last few messages* before the error appeared? [image: aiderdeskerror2.png]

…

On Wed, May 13, 2026 at 7:06 PM Vladimir Hrusovsky ***@***.***> wrote: *wladimiiir* left a comment (hotovo/aider-desk#785) <#785 (comment)> This error (Assistant response prefill is incompatible with enable_thinking) does not originate from AiderDesk — it is returned by the *llama.cpp inference engine* that powers your Jan.ai backend. Jan.ai uses llama.cpp (via cortex.llamacpp) under the hood for serving GGUF models. The llama.cpp server enforces a constraint: when enable_thinking is active (which is the case for thinking-capable models like Qwen 3.5/3.6), the last message in the API request cannot have role: "assistant" (known as "assistant response prefill"). This can get triggered during multi-turn tool-use conversations where the message history naturally ends with an assistant message. This is a known upstream issue — see ggml-org/llama.cpp#20861 <ggml-org/llama.cpp#20861> and mastra-ai/mastra#15234 <mastra-ai/mastra#15234> for the same error with Qwen models. To help me investigate further: 1. *Which mode were you using?* — Agent mode, or one of the Aider modes (Code, Ask, Architect, etc.)? 2. *Does re-sending the same prompt fix it?* If it's intermittent, it may be triggered by specific conversation patterns (e.g., after tool calls). 3. Could you share the *context.json* file from the project's .aider-desk/tasks/{taskId} directory? This would show me the exact message sequence being sent. (You can sanitize any sensitive content.) 4. What were the *last few messages* before the error appeared? (e.g., was it after a tool call, after a long conversation, right from the start, etc.) As a potential workaround, you could try disabling thinking/reasoning mode in Jan.ai's model settings if available, or check if your Jan.ai version exposes any llama.cpp flags related to prefill-assistant. — Reply to this email directly, view it on GitHub <#785 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHQVDIQ4RVYP74SRSDPYMWL42SMPHAVCNFSM6AAAAACY263ZBGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DINBSHE2DGMBYGQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[akierum]

I downgraded to v0.64.0

and now no more error. And no more qwen 3.6 27b q8 LLM stuck in loops either.

[akierum]

did you get rar file I sent you by email with requested files?

[wladimiiir]

No, the email is for notifications only, it does not accept the files. You will need to upload them here or send me a link.

[akierum]

Hello, I would like to provide it in confidential way, these messages are posted on github.

…

On Tue, May 19, 2026 at 8:32 AM Vladimir Hrusovsky ***@***.***> wrote: *wladimiiir* left a comment (hotovo/aider-desk#785) <#785 (comment)> No, the email is for notifications only, it does not accept the files. You will need to upload them here or send me a link. — Reply to this email directly, view it on GitHub <#785 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHQVDIQUP4XBB73BDOPFLAD43PWXPAVCNFSM6AAAAACY263ZBGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DIOBUG4ZDCMZUHE> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[wladimiiir]

@akierum
Feel free to DM me on Discord.