diff --git a/memento-note/lib/ai/providers/custom-openai.ts b/memento-note/lib/ai/providers/custom-openai.ts
index fcf3007..63914c9 100644
--- a/memento-note/lib/ai/providers/custom-openai.ts
+++ b/memento-note/lib/ai/providers/custom-openai.ts
@@ -119,25 +119,40 @@ export class CustomOpenAIProvider implements AIProvider {
 
   async generateWithTools(options: ToolUseOptions): Promise<ToolCallResult> {
     const { tools, maxSteps = 10, systemPrompt, messages, prompt } = options
-    const opts: Record<string, any> = {
-      model: this.model,
-      tools,
-      stopWhen: stepCountIs(maxSteps),
-    }
-    if (systemPrompt) opts.system = systemPrompt
-    if (messages) opts.messages = messages
-    else if (prompt) opts.prompt = prompt
 
-    const result = await aiGenerateText(opts as any)
-    return {
-      toolCalls: result.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
-      toolResults: result.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
-      text: result.text,
-      steps: result.steps?.map((step: any) => ({
+    const buildOpts = (steps: number): Record<string, any> => {
+      const opts: Record<string, any> = { model: this.model, tools, stopWhen: stepCountIs(steps) }
+      if (systemPrompt) opts.system = systemPrompt
+      if (messages) opts.messages = messages
+      else if (prompt) opts.prompt = prompt
+      return opts
+    }
+
+    const toResult = (r: any): ToolCallResult => ({
+      toolCalls: r.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
+      toolResults: r.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
+      text: r.text,
+      steps: r.steps?.map((step: any) => ({
         text: step.text,
         toolCalls: step.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
-        toolResults: step.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || []
-      })) || []
+        toolResults: step.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
+      })) || [],
+    })
+
+    try {
+      const result = await aiGenerateText(buildOpts(maxSteps) as any)
+      return toResult(result)
+    } catch (err: any) {
+      // DeepSeek reasoning/thinking models require reasoning_content to be passed back
+      // between multi-step calls, which the AI SDK doesn't handle via the OpenAI-compat layer.
+      // Retry with a single step so the model calls the tool directly.
+      const msg: string = err?.message || String(err)
+      if (msg.includes('reasoning_content') || msg.includes('thinking mode')) {
+        console.warn('[CustomOpenAI] Reasoning model detected — retrying with maxSteps=1')
+        const result = await aiGenerateText(buildOpts(1) as any)
+        return toResult(result)
+      }
+      throw err
     }
   }
 
diff --git a/memento-note/lib/ai/providers/deepseek.ts b/memento-note/lib/ai/providers/deepseek.ts
index 1e1b86d..b59897b 100644
--- a/memento-note/lib/ai/providers/deepseek.ts
+++ b/memento-note/lib/ai/providers/deepseek.ts
@@ -103,25 +103,40 @@ export class DeepSeekProvider implements AIProvider {
 
   async generateWithTools(options: ToolUseOptions): Promise<ToolCallResult> {
     const { tools, maxSteps = 10, systemPrompt, messages, prompt } = options
-    const opts: Record<string, any> = {
-      model: this.model,
-      tools,
-      stopWhen: stepCountIs(maxSteps),
-    }
-    if (systemPrompt) opts.system = systemPrompt
-    if (messages) opts.messages = messages
-    else if (prompt) opts.prompt = prompt
 
-    const result = await aiGenerateText(opts as any)
-    return {
-      toolCalls: result.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
-      toolResults: result.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
-      text: result.text,
-      steps: result.steps?.map((step: any) => ({
+    const buildOpts = (steps: number): Record<string, any> => {
+      const opts: Record<string, any> = { model: this.model, tools, stopWhen: stepCountIs(steps) }
+      if (systemPrompt) opts.system = systemPrompt
+      if (messages) opts.messages = messages
+      else if (prompt) opts.prompt = prompt
+      return opts
+    }
+
+    const toResult = (r: any): ToolCallResult => ({
+      toolCalls: r.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
+      toolResults: r.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
+      text: r.text,
+      steps: r.steps?.map((step: any) => ({
         text: step.text,
         toolCalls: step.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
-        toolResults: step.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || []
-      })) || []
+        toolResults: step.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
+      })) || [],
+    })
+
+    try {
+      const result = await aiGenerateText(buildOpts(maxSteps) as any)
+      return toResult(result)
+    } catch (err: any) {
+      // DeepSeek reasoning/thinking models require reasoning_content to be passed back
+      // between multi-step calls, which the AI SDK doesn't handle automatically.
+      // Retry with a single step so the model calls the tool directly without multi-turn.
+      const msg: string = err?.message || String(err)
+      if (msg.includes('reasoning_content') || msg.includes('thinking mode')) {
+        console.warn('[DeepSeek] Reasoning model detected — retrying with maxSteps=1')
+        const result = await aiGenerateText(buildOpts(1) as any)
+        return toResult(result)
+      }
+      throw err
     }
   }