diff --git a/packages/opencode/src/provider/provider.ts b/packages/opencode/src/provider/provider.ts
index 9aa1b6304c12..830f5da1a883 100644
--- a/packages/opencode/src/provider/provider.ts
+++ b/packages/opencode/src/provider/provider.ts
@@ -1006,7 +1006,7 @@ function fromModelsDevModel(provider: ModelsDev.Provider, model: ModelsDev.Model
         video: model.modalities?.output?.includes("video") ?? false,
         pdf: model.modalities?.output?.includes("pdf") ?? false,
       },
-      interleaved: model.interleaved ?? false,
+      interleaved: model.interleaved ?? (model.reasoning ? { field: "reasoning_content" } : false),
     },
     release_date: model.release_date ?? "",
     variants: {},
@@ -1177,7 +1177,7 @@ const layer: Layer.Layer<
                     model.modalities?.output?.includes("video") ?? existingModel?.capabilities.output.video ?? false,
                   pdf: model.modalities?.output?.includes("pdf") ?? existingModel?.capabilities.output.pdf ?? false,
                 },
-                interleaved: model.interleaved ?? existingModel?.capabilities.interleaved ?? false,
+                interleaved: model.interleaved ?? existingModel?.capabilities.interleaved ?? (model.reasoning ? { field: "reasoning_content" } : false),
               },
               cost: {
                 input: model?.cost?.input ?? existingModel?.cost?.input ?? 0,
diff --git a/packages/opencode/src/provider/transform.ts b/packages/opencode/src/provider/transform.ts
index 67b02c089602..05b352da5a81 100644
--- a/packages/opencode/src/provider/transform.ts
+++ b/packages/opencode/src/provider/transform.ts
@@ -176,7 +176,8 @@ function normalizeMessages(
   }
 
   // Deepseek requires all assistant messages to have reasoning on them
-  if (model.api.id.includes("deepseek")) {
+  // Check both API ID and model ID to cover OpenRouter-routed DeepSeek models
+  if (model.api.id.includes("deepseek") || model.id.includes("deepseek")) {
     msgs = msgs.map((msg) => {
       if (msg.role !== "assistant") return msg
       if (Array.isArray(msg.content)) {
@@ -195,6 +196,7 @@ function normalizeMessages(
 
   if (typeof model.capabilities.interleaved === "object" && model.capabilities.interleaved.field) {
     const field = model.capabilities.interleaved.field
+    const sdk = sdkKey(model.api.npm) ?? "openaiCompatible"
     return msgs.map((msg) => {
       if (msg.role === "assistant" && Array.isArray(msg.content)) {
         const reasoningParts = msg.content.filter((part: any) => part.type === "reasoning")
@@ -203,17 +205,21 @@ function normalizeMessages(
         // Filter out reasoning parts from content
         const filteredContent = msg.content.filter((part: any) => part.type !== "reasoning")
 
-        // Include reasoning_content | reasoning_details directly on the message for all assistant messages.
-        // Always set the field even when empty — some providers (e.g. DeepSeek) may return empty
-        // reasoning_content which still needs to be sent back in subsequent requests.
+        // Preserve existing providerOptions[field] when content no longer has reasoning parts
+        // (e.g. after a prior transform pass already extracted them). The @ai-sdk yG converter
+        // resolves reasoning_content from providerOptions → overwriting it with empty string
+        // on the second pass causes DeepSeek 400: "reasoning_content must be passed back".
+        const existingField = msg.providerOptions?.[sdk]?.[field]
+        const resolvedText = reasoningText || existingField || ""
+
         return {
           ...msg,
           content: filteredContent,
           providerOptions: {
             ...msg.providerOptions,
-            openaiCompatible: {
-              ...msg.providerOptions?.openaiCompatible,
-              [field]: reasoningText,
+            [sdk]: {
+              ...msg.providerOptions?.[sdk],
+              [field]: resolvedText,
             },
           },
         }
@@ -223,6 +229,44 @@ function normalizeMessages(
     })
   }
 
+  // When reasoning is active but interleaved is not configured, still inject empty reasoning_content
+  // for ALL assistant messages. This covers historical messages from DB that were stored before
+  // reasoning mode was enabled — they have no reasoning part to extract but DeepSeek's API still
+  // requires reasoning_content on every assistant turn in thinking mode.
+  if (model.capabilities.reasoning) {
+    msgs = msgs.map((msg) => {
+      if (msg.role !== "assistant") return msg
+      if (Array.isArray(msg.content)) {
+        const sdk = sdkKey(model.api.npm) ?? "openaiCompatible"
+        return {
+          ...msg,
+          providerOptions: {
+            ...msg.providerOptions,
+            [sdk]: {
+              ...msg.providerOptions?.[sdk],
+              reasoning_content: "",
+            },
+          },
+        }
+      }
+      if (typeof msg.content === "string") {
+        const sdk = sdkKey(model.api.npm) ?? "openaiCompatible"
+        return {
+          ...msg,
+          content: [{ type: "text" as const, text: msg.content }, { type: "reasoning" as const, text: "" }],
+          providerOptions: {
+            ...msg.providerOptions,
+            [sdk]: {
+              ...msg.providerOptions?.[sdk],
+              reasoning_content: "",
+            },
+          },
+        }
+      }
+      return msg
+    })
+  }
+
   return msgs
 }