When using the qwen3.5-122b-a10b model (via nvidia provider), the following error is returned: System message must be at the beginning #16560

Activity

1. fanyapeng

   added

   bugSomething isn't workingSomething isn't working

   on Mar 8, 2026
2. github-actions

   assigned

   thdxr

   on Mar 8, 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 Mar 8, 2026
4. github-actions commented on Mar 8, 2026

   

   github-actionsBot

   on Mar 8, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • Duplicate System Message Sent #1301: Duplicate System Message Sent — OpenCode sends multiple system messages; certain models (like the NVIDIA-hosted Qwen model) reject requests where the system message is not at the very beginning or is duplicated.
   • Ministral 3 support: error "Only user, assistant and tool roles are supported, got system" #5034: Ministral 3 support — same root cause, a different model rejecting system messages that are not at the beginning of the conversation.

   If your issue is the same underlying problem, consider adding a comment or reaction there rather than opening a new report.
5. paulgear commented on Mar 8, 2026

   

   paulgear

   on Mar 8, 2026

   More actions

   This is a dup of #5034 (and possibly a recurrence of #1301), and I've produced a PR #15018 to fix it. Please add your upvote so it can get merged.

   👍React with 👍15Reacted by cshuai, Yichao Li, kmtbang, Lu Lu, Luong Hoang Minh, wangzhipeng0916, ice5631024721, kevin, 宫丞施, apnea and 5 more
6. brainwang commented on Apr 10, 2026

   

   brainwang

   on Apr 10, 2026

   More actions

   Any workaround?
7. sgmihai commented on Apr 16, 2026

   

   sgmihai

   on Apr 16, 2026

   More actions

   qwen3.6 32b in vllm, getting this message in opencode + omo slim, every now and then
8. thdxr

   assigned

   kitlangton

   and unassigned

   thdxr

   on Apr 26, 2026
9. thdxr

   removed

   bugSomething isn't workingSomething isn't working

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

   on May 3, 2026
10. robertoberto commented on May 23, 2026

    

    robertoberto

    on May 23, 2026

    More actions

    same bug as #15059. @paulgear's #15018 was the right shape, and #16981 (currently open and awaiting code-owner review) generalizes it with a capability flag.

    still broken on opencode v1.14.39 against Qwen/Qwen3.6-27B and Qwen/Qwen3.6-35B-A3B on vLLM v0.21.0, even with no plugins — the title-generator path produces [system, system, user] on its own. running the equivalent of #15018 locally resolves the 400s without affecting any other provider.

    diagnosis + repro in #15059.
11. github-actions commented on Jul 23, 2026

    

    github-actionsBot

    on Jul 23, 2026 – with GitHub Actions

    Contributor

    More actions

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

    closed this as not plannedon Jul 23, 2026
13. brusacco commented on Aug 16, 2026

    

    brusacco

    on Aug 16, 2026

    More actions

    Build a little proxy in Node to fix this issue; this is happening now on Qwen 3.8

    `const http = require("http");

    const PORT = 11435;
    const OLLAMA_HOST = "192.168.0.32";
    const OLLAMA_PORT = 11434;

    /**

    • Qwen 3.8 requires system messages to appear only at the beginning.
    • OpenCode may send multiple system messages throughout the message array.
    • This function:
    • 1. Collects every system message.
    • 2. Merges their content.
    • 3. Places one system message at position 0.
    • 4. Preserves every non-system message unchanged.
         */
         function mergeSystemMessages(messages) {
         if (!Array.isArray(messages)) {
         return messages;
         }

    const systemMessages = [];
    const otherMessages = [];
    
    for (const message of messages) {
        if (message.role === "system") {
            systemMessages.push(message);
        } else {
            otherMessages.push(message);
        }
    }
    
    if (systemMessages.length === 0) {
        return messages;
    }
    
    // Preserve string or structured content.
    const allStrings = systemMessages.every(
        (message) => typeof message.content === "string"
    );
    
    let content;
    
    if (allStrings) {
        content = systemMessages
            .map((message) => message.content)
            .filter(Boolean)
            .join("\n\n");
    } else {
        // Handle OpenAI-style structured content arrays.
        content = systemMessages.flatMap((message) => {
            if (Array.isArray(message.content)) {
                return message.content;
            }
    
            if (typeof message.content === "string") {
                return [
                    {
                        type: "text",
                        text: message.content,
                    },
                ];
            }
    
            return [];
        });
    }
    
    console.log(
        `[proxy] merged ${systemMessages.length} system messages`
    );
    
    return [
        {
            role: "system",
            content,
        },
        ...otherMessages,
    ];
    

    }

    const server = http.createServer((req, res) => {
    // We only need to modify chat completions.
    if (
    req.method === "POST" &&
    req.url.startsWith("/v1/chat/completions")
    ) {
    let body = "";

        req.on("data", (chunk) => {
            body += chunk;
        });
    
        req.on("end", () => {
            try {
                const payload = JSON.parse(body);
    
                if (Array.isArray(payload.messages)) {
                    console.log(
                        `[proxy] ${payload.model} - ${payload.messages.length} messages`
                    );
    
                    console.log(
                        "[proxy] roles:",
                        payload.messages.map((m) => m.role).join(" → ")
                    );
    
                    payload.messages = mergeSystemMessages(payload.messages);
    
                    console.log(
                        "[proxy] forwarded roles:",
                        payload.messages.map((m) => m.role).join(" → ")
                    );
                }
    
                const modifiedBody = JSON.stringify(payload);
    
                const headers = {
                    ...req.headers,
                    host: `${OLLAMA_HOST}:${OLLAMA_PORT}`,
                    "content-length": Buffer.byteLength(modifiedBody),
                };
    
                const proxyReq = http.request(
                    {
                        hostname: OLLAMA_HOST,
                        port: OLLAMA_PORT,
                        path: req.url,
                        method: req.method,
                        headers,
                    },
                    (proxyRes) => {
                        res.writeHead(proxyRes.statusCode, proxyRes.headers);
    
                        // Important:
                        // pipe directly so Ollama SSE streaming remains streaming.
                        proxyRes.pipe(res);
                    }
                );
    
                proxyReq.on("error", (error) => {
                    console.error("[proxy] Ollama error:", error);
    
                    if (!res.headersSent) {
                        res.writeHead(502, {
                            "Content-Type": "application/json",
                        });
                    }
    
                    res.end(
                        JSON.stringify({
                            error: {
                                message: error.message,
                                type: "proxy_error",
                            },
                        })
                    );
                });
    
                proxyReq.write(modifiedBody);
                proxyReq.end();
            } catch (error) {
                console.error("[proxy] Invalid request:", error);
    
                res.writeHead(400, {
                    "Content-Type": "application/json",
                });
    
                res.end(
                    JSON.stringify({
                        error: {
                            message: error.message,
                            type: "proxy_error",
                        },
                    })
                );
            }
        });
    
        return;
    }
    
    /*
     * Everything else is passed through untouched.
     *
     * This lets things such as /v1/models continue working.
     */
    
    const proxyReq = http.request(
        {
            hostname: OLLAMA_HOST,
            port: OLLAMA_PORT,
            path: req.url,
            method: req.method,
            headers: {
                ...req.headers,
                host: `${OLLAMA_HOST}:${OLLAMA_PORT}`,
            },
        },
        (proxyRes) => {
            res.writeHead(proxyRes.statusCode, proxyRes.headers);
            proxyRes.pipe(res);
        }
    );
    
    proxyReq.on("error", (error) => {
        console.error("[proxy] Ollama error:", error);
    
        res.writeHead(502, {
            "Content-Type": "application/json",
        });
    
        res.end(
            JSON.stringify({
                error: {
                    message: error.message,
                    type: "proxy_error",
                },
            })
        );
    });
    
    req.pipe(proxyReq);
    

    });

    server.listen(PORT, "0.0.0.0", () => {
    console.log(Qwen/OpenCode proxy listening on :${PORT});
    console.log(
    Forwarding to http://${OLLAMA_HOST}:${OLLAMA_PORT}
    );
    });`