Ministral 3 support: error "Only user, assistant and tool roles are supported, got system"

Description

I wish to use opencode with Ministral 3 14B which was recently released. However, opencode raises an error immediately:

Only user, assistant and tool roles are supported, got system

To understand it better, here is first my setup, and how to reproduce the error, which I expect will occur even for the Ministral 3 3B, as we will see later.

I run the server using weights downloaded from unsloth:

huggingface-cli download unsloth/Ministral-3-14B-Reasoning-2512-GGUF Ministral-3-14B-Reasoning-2512-Q5_K_M.gguf
docker run --gpus all -v /data/ml/models/gguf:/models -p 11432:8080 ghcr.io/ggml-org/llama.cpp:full-cuda -s --host 0.0.0.0 -m "/models/Ministral-3-14B-Reasoning-2512-Q5_K_M.gguf" --ctx-size 0 -ub 2048 -b 2048 --n-gpu-layers 99 -fa on -c 65536

I then run opencode with the following configuration, as per the documentation:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "llama-server (local)",
      "options": {
        "baseURL": "http://127.0.0.1:11432/v1"
      },
      "models": {
        "ministral-3-14b": {
          "name": "Ministral 3 14B (local)"
        }
      }
    }
  }
}

Then I request anything, and get the error pasted at the top.

The reason I get this error is that llama.cpp relies on the Jinja template inset in the GGUF, which it copied from the original Ministral repository. It loops over loop_messages, expecting not to have system messages. It only expects a system message on the first message.

However, that clearly is not the case with opencode, which can seemingly have up to 3 system messages: https://github.com/sst/opencode/blob/ee4437ff32fc2acbd2220060fc980a096730bcee/packages/opencode/src/session/prompt.ts#L651-L660

For opencode to support Ministral, it needs a single system prompt with the current Jinja template, concatenating them.

OpenCode version

1.0.129

Steps to reproduce

I run the server using weights downloaded from unsloth:

huggingface-cli download unsloth/Ministral-3-14B-Reasoning-2512-GGUF Ministral-3-14B-Reasoning-2512-Q5_K_M.gguf
docker run --gpus all -v /data/ml/models/gguf:/models -p 11432:8080 ghcr.io/ggml-org/llama.cpp:full-cuda -s --host 0.0.0.0 -m "/models/Ministral-3-14B-Reasoning-2512-Q5_K_M.gguf" --ctx-size 0 -ub 2048 -b 2048 --n-gpu-layers 99 -fa on -c 65536

I then run opencode with the following configuration, as per the documentation:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "llama-server (local)",
      "options": {
        "baseURL": "http://127.0.0.1:11432/v1"
      },
      "models": {
        "ministral-3-14b": {
          "name": "Ministral 3 14B (local)"
        }
      }
    }
  }
}

Then I enter a message, and I see the error:

Only user, assistant and tool roles are supported, got system

Screenshot and/or share link

Operating System

Ubuntu 24.04.2 LTS

Terminal

Alacritty

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

BUG: TemplateError #2272: TemplateError about conversation roles needing to alternate - similar root cause where consecutive system messages are sent to servers with Jinja templates
OpenCode sends tools (and Jinja tool template) to llama.cpp results in 500 error unless --jinja; with --jinja, template crashes (reject filter) #1890: OpenCode sends tools and Jinja tool template to llama.cpp causing 500 errors - addresses llama.cpp compatibility issues with multiple system messages
Input text displays in terminal instead of input box, Enter creates newline instead of submitting (V 1.0.20) #4107: Databricks Gemini 2.5 Pro error "Gemini models only support one system prompt" - indicates the codebase sends multiple system prompts, which some providers reject

All three issues stem from the same underlying problem: OpenCode generates up to 3 system messages (as referenced in prompt.ts:651-660), but many model providers and Jinja templates expect only a single system message, especially at the start of conversation.

Feel free to ignore if your case differs, but these likely share the same root cause and solution.

This one might be related to ggml-org/llama.cpp#17700 and ggml-org/llama.cpp#17713.

EDIT: actually I am no longer able to reproduce this:

The error message changes to
After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results
when I apply ggml-org/llama.cpp#17713.

I think this is an issue on llama.cpp side right

Well, I don't know. The problem is in default ministral jinja template https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512/blob/f8aa436a36b5bf3c59e3c2e8b19368de9cf69914/chat_template.jinja, which basically requires alternating roles in messages. Therefore:

{
    ...
    "messages": [
        {
            "role":"system",
            "content":"You are"
        },
        {
            "role":"system",
            "content":"Here is"
        },
        {
            "role":"user",
            "content":"test"
        }
    ],
    ...
}

and

{
    ...
    "messages": [
        {
            "role":"system",
            "content":"You are"
        },
        {
            "role":"user",
            "content":"test 1"
        }
        {
            "role":"user",
            "content":"test 2"
        }
    ],
    ...
}

are both problematic. The first snippet is what opencode basically sends by default and the template throws an error: Only user, assistant and tool roles are supported, got system. at row 124, column 111... for it. The issue seems to be related to #2272.

Still not working, everything is fine on the llame.cpp side since they today but it's still not working with opencode :( and devstrall too ....Is there a quick bypass or fix ?

For quick fix you can merge all system prompts into one here, or modify mistral chat template and pass it via --chat-template. Would be good if opencode allowed to customize this.

A workaround is to use Mistral Vibe, but I hope opencode will support those models too

I think this should be fixed? The system prompts should be sent as merged (only exception currently is if u are using anthropic)

I'm getting this issue., with opencode 1.0.218 using devstral-2-small.

$ llama-server --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 3090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
version: 1087 (66d767f)
built with GNU 15.2.1 for Linux x86_64

llama-server \
    -hf unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Q3_K_S \
    --jinja -ngl 99 --threads -1 --ctx-size 16384 \
    --temp 0.15 \
                --no-mmap

errors from llama-server

srv operator(): got exception: {"error":{"code":500,"message":"After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results. at row 189, column 164:\n {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}\n
{{- raise_exception('After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results.') }}\n ^\n
{%- endif %}\n at row 189, column 13:\n {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}\n {{- raise_exception('After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results.') }}\n ^\n {%- endif %}\n at row 188, column 69:\n {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}\n {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}\n
^\n {{- raise_exception('After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results.') }}\n at row 188, column 9:\n {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}\n {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}\n ^\n
{{- raise_exception('After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results.') }}\n at row 187, column 176:\n{%- for message in loop_messages %}\n {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}\n ^\n
{%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}\n at row 187, column 5:\n{%- for message in loop_messages %}\n {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}\n ^\n {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}\n at row 186, column 36:\n{%- set ns = namespace(index=0) %}\n{%- for message in loop_messages %}\n ^
\n {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}\n at row 186, column 1:\n{%- set ns = namespace(index=0) %}\n{%- for message in loop_messages %}\n^\n {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}\n at row 1, column 30:\n{#- Unsloth template fixes #}\n
^\n{%- set yesterday_day = strftime_now("%d") %}\n","type":"server_error"}}

(didn't put it in a code block as it tried to make it one really long line)

devstral 2 also affected :/

Well, I don't know. The problem is in default ministral jinja template

I think it's desirable that all models would support non-alternating interactions, but that's probably an assumption too far on opencode's part, since the models are usually trained for back-and-forth dialog. Evidence would be that mistral vibe works, opencode does not.

A quick fix might be to join contiguous responses from either party.

It looks like opencode constructs 2 system messages by default, and qwen 3.5's default template expects 1 system message at the beginning only.

opencode/packages/opencode/src/session/llm.ts

Lines 83 to 93 in f2100dc

    
           await Plugin.trigger( 
        
             "experimental.chat.system.transform", 
        
             { sessionID: input.sessionID, model: input.model }, 
        
             { system }, 
        
           ) 
        
           // rejoin to maintain 2-part structure for caching if header unchanged 
        
           if (system.length > 2 && system[0] === header) { 
        
             const rest = system.slice(1) 
        
             system.length = 0 
        
             system.push(header, rest.join("\n")) 
        
           }

It looks like opencode constructs 2 system messages by default, and qwen 3.5's default template only expects 1 system message at the beginning only.

I made a fix for this at #15018 - maybe if we can get some 👍 on it it might get merged.

@paulgear

Thanks for this @paulgear - I built with your fix and I can finally use opencode again with the new Qwen 3.5 models.

Would love to run a release from anomalyco, though. Will keep an eye on this to see when your fix gets merged.

Thanks again!

For non-anthropic providers, combine system prompts

Reproducing with:

OpenCode v1.2.24/1.2.25
vllm-mlx v0.2.6 serving Qwen3.5-27B-8bit (MLX quantized)
macOS Darwin 25.3.0, Apple Silicon M2 Ultra

The root cause is that OpenCode sends multiple system messages in its API requests. The Qwen 3.5 official chat template (chat_template.jinja shipped with the model weights) explicitly rejects this at line 84-85:

{%- if message.role == "system" %}
    {%- if not loop.first %}
        {{- raise_exception('System message must be at the beginning.') }}
    {%- endif %}

The template only allows 4 roles: system, user, assistant, tool. Any other role (like developer) hits the catch-all at line 143-144:

{%- else %}
    {{- raise_exception('Unexpected message role.') }}

This affects ALL Qwen 3.5 models (27B, 35B, 122B) across ALL servers that use the model's built-in template (vllm-mlx, vllm, llama.cpp without custom template override).

Server-side workaround (custom template) exists but shouldn't be necessary — OpenCode should consolidate system messages for non-Anthropic providers as PR #15018 proposes.

+1 for PR #15018 — this is a critical blocker for Qwen 3.5 users.

Maybe the new issue should be created because it’s not just about Ministral, it’s a much broader problem.

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.

Navigation Menu

Repository navigation

Ministral 3 support: error "Only user, assistant and tool roles are supported, got system" #5034

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

7 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	await Plugin.trigger(
	"experimental.chat.system.transform",
	{ sessionID: input.sessionID, model: input.model },
	{ system },
	)
	// rejoin to maintain 2-part structure for caching if header unchanged
	if (system.length > 2 && system[0] === header) {
	const rest = system.slice(1)
	system.length = 0
	system.push(header, rest.join("\n"))
	}

Ministral 3 support: error "Only user, assistant and tool roles are supported, got system" #5034

Description

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Dec 4, 2025

arkamar commented on Dec 5, 2025

rekram1-node commented on Dec 7, 2025

arkamar commented on Dec 7, 2025

bamer commented on Dec 10, 2025

kurnevsky commented on Dec 10, 2025

grigio commented on Dec 12, 2025

rekram1-node commented on Dec 16, 2025

haydonryan commented on Dec 30, 2025

lee-b commented on Jan 27, 2026

lee-b commented on Jan 27, 2026

7 remaining items

cyangle commented on Feb 27, 2026

paulgear commented on Feb 27, 2026

0xDEADBEEF-td commented on Mar 3, 2026

Thump604 commented on Mar 13, 2026

jacekpoplawski commented on Mar 13, 2026

github-actions commented on Jul 3, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions