fix: handle system messages for non-Anthropic providers

- Add systemMessage capability schema ('single' | 'multiple')
- Add getSystemMessageMode helper with provider-aware defaults
- Anthropic keeps multiple system messages; all others merge into one
- Add reasoning_content support for DeepSeek/oMLX models
- Merge system messages in convert layer for OpenAI-compatible APIs
- Handle lone system message edge case (convert to user message)

Fixes 'Chat template error' for Qwen, Llama, Ollama, etc.

Closes #5034

Author: jwcrystal

packages/opencode/src/provider/provider.ts

@@ -845,6 +845,7 @@ const ProviderCapabilities = Schema.Struct({
   input: ProviderModalities,
   output: ProviderModalities,
   interleaved: ProviderInterleaved,
+  systemMessage: Schema.Union([Schema.Literal("single"), Schema.Literal("multiple")]).optional(),
 })
 
 const ProviderCacheCost = Schema.Struct({
@@ -1007,6 +1008,7 @@ function fromModelsDevModel(provider: ModelsDev.Provider, model: ModelsDev.Model
         pdf: model.modalities?.output?.includes("pdf") ?? false,
       },
       interleaved: model.interleaved ?? false,
+      systemMessage: undefined,
     },
     release_date: model.release_date ?? "",
     variants: {},
@@ -1171,7 +1173,7 @@ const layer: Layer.Layer<
                   pdf: model.modalities?.input?.includes("pdf") ?? existingModel?.capabilities.input.pdf ?? false,
                 },
                 output: {
-                  text: model.modalities?.output?.includes("text") ?? existingModel?.capabilities.output.text ?? true,
+                  text: model.modalities?.output?.includes("text") ?? true,
                   audio:
                     model.modalities?.output?.includes("audio") ?? existingModel?.capabilities.output.audio ?? false,
                   image:
@@ -1181,6 +1183,7 @@ const layer: Layer.Layer<
                   pdf: model.modalities?.output?.includes("pdf") ?? existingModel?.capabilities.output.pdf ?? false,
                 },
                 interleaved: model.interleaved ?? false,
+                systemMessage: undefined,
               },
               cost: {
                 input: model?.cost?.input ?? existingModel?.cost?.input ?? 0,

packages/opencode/src/provider/sdk/copilot/chat/convert-to-openai-compatible-chat-messages.ts

@@ -12,15 +12,35 @@ function getOpenAIMetadata(message: { providerOptions?: SharedV3ProviderOptions
 
 export function convertToOpenAICompatibleChatMessages(prompt: LanguageModelV3Prompt): OpenAICompatibleChatPrompt {
   const messages: OpenAICompatibleChatPrompt = []
+
+  const systemPrompt: string[] = []
+  for (const { role, content } of prompt) {
+    if (role === "system") {
+      systemPrompt.push(content)
+    }
+  }
+
+  const hasSystem = systemPrompt.length > 0
+  const hasOthers = prompt.some((m) => m.role !== "system")
+
+  if (hasSystem) {
+    if (hasOthers) {
+      messages.push({
+        role: "system",
+        content: systemPrompt.join("\n\n"),
+      })
+    } else {
+      messages.push({
+        role: "user",
+        content: systemPrompt.join("\n\n"),
+      })
+    }
+  }
+
   for (const { role, content, ...message } of prompt) {
     const metadata = getOpenAIMetadata({ ...message })
     switch (role) {
       case "system": {
-        messages.push({
-          role: "system",
-          content: content,
-          ...metadata,
-        })
         break
       }
 

packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts

@@ -226,8 +226,8 @@ export class OpenAICompatibleChatLanguageModel implements LanguageModelV3 {
       })
     }
 
-    // reasoning content (Copilot uses reasoning_text):
-    const reasoning = choice.message.reasoning_text
+    // reasoning content (Copilot uses reasoning_text, DeepSeek/oMLX uses reasoning_content):
+    const reasoning = choice.message.reasoning_text ?? choice.message.reasoning_content
     if (reasoning != null && reasoning.length > 0) {
       content.push({
         type: "reasoning",
@@ -477,8 +477,8 @@ export class OpenAICompatibleChatLanguageModel implements LanguageModelV3 {
               reasoningOpaque = delta.reasoning_opaque
             }
 
-            // enqueue reasoning before text deltas (Copilot uses reasoning_text):
-            const reasoningContent = delta.reasoning_text
+            // enqueue reasoning before text deltas (Copilot uses reasoning_text, DeepSeek/oMLX uses reasoning_content):
+            const reasoningContent = delta.reasoning_text ?? delta.reasoning_content
             if (reasoningContent) {
               if (!isActiveReasoning) {
                 controller.enqueue({
@@ -757,6 +757,7 @@ const OpenAICompatibleChatResponseSchema = z.object({
         // Copilot-specific reasoning fields
         reasoning_text: z.string().nullish(),
         reasoning_opaque: z.string().nullish(),
+        reasoning_content: z.string().nullish(),
         tool_calls: z
           .array(
             z.object({
@@ -792,6 +793,7 @@ const createOpenAICompatibleChatChunkSchema = <ERROR_SCHEMA extends z.core.$ZodT
               // Copilot-specific reasoning fields
               reasoning_text: z.string().nullish(),
               reasoning_opaque: z.string().nullish(),
+              reasoning_content: z.string().nullish(),
               tool_calls: z
                 .array(
                   z.object({

packages/opencode/src/session/llm.ts

@@ -29,6 +29,23 @@ const log = Log.create({ service: "llm" })
 export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
 type Result = Awaited<ReturnType<typeof streamText>>
 
+function getSystemMessageMode(model: Provider.Model): "single" | "multiple" {
+  if (model.capabilities.systemMessage) {
+    return model.capabilities.systemMessage
+  }
+  const providerDefaults: Record<string, "single" | "multiple"> = {
+    anthropic: "multiple",
+  }
+  const mode = providerDefaults[model.providerID] ?? "single"
+  if (!providerDefaults[model.providerID]) {
+    log.info("Using default 'single' systemMessage mode for provider", {
+      providerID: model.providerID,
+      modelID: model.id,
+    })
+  }
+  return mode
+}
+
 export type StreamInput = {
   user: MessageV2.User
   sessionID: string
@@ -150,12 +167,14 @@ const live: Layer.Layer<
         : isWorkflow
           ? input.messages
           : [
-              ...system.map(
-                (x): ModelMessage => ({
-                  role: "system",
-                  content: x,
-                }),
-              ),
+              ...(getSystemMessageMode(input.model) === "multiple"
+                ? system.map(
+                    (x): ModelMessage => ({
+                      role: "system",
+                      content: x,
+                    }),
+                  )
+                : ([{ role: "system", content: system.join("\n") }] as ModelMessage[])),
               ...input.messages,
             ]