Fix studio crash on Mac by deferring unsloth_zoo import

Move `from unsloth_zoo.rl_environments import check_signal_escape_patterns`
from module-level to inside `_check_code_safety()` and wrap it in a
try/except for NotImplementedError. The top-level import triggers
`unsloth_zoo.__init__` which calls `get_device_type()` at module scope,
raising NotImplementedError on Apple Silicon. On Mac the safety check
gracefully skips since unsloth_zoo is unavailable, while GGUF inference
via llama.cpp continues to work.

studio/backend/core/inference/tools.py

@@ -17,7 +17,6 @@ import tempfile
 import threading
 
 from loggers import get_logger
-from unsloth_zoo.rl_environments import check_signal_escape_patterns
 
 logger = get_logger(__name__)
 
@@ -171,6 +170,11 @@ def _check_code_safety(code: str) -> str | None:
     Returns an error message string if the code is unsafe, or None if OK.
     """
     # Check for signal/timeout escape patterns
+    try:
+        from unsloth_zoo.rl_environments import check_signal_escape_patterns
+    except (ImportError, NotImplementedError):
+        # unsloth_zoo may not be available or may fail on non-GPU platforms (e.g. Mac)
+        return None
     safe, info = check_signal_escape_patterns(code)
     if not safe:
         reasons = [

studio/frontend/src/components/assistant-ui/tool-ui-python.tsx

@@ -17,7 +17,7 @@ import {
 
 const MAX_DISPLAY = 10_000;
 const COPY_RESET_MS = 2000;
-const SHIKI_THEME = ["github-light", "github-dark"] as const;
+const SHIKI_THEME = ["github-light", "github-dark"] as ["github-light", "github-dark"];
 
 function truncate(text: string): string {
   return text.length <= MAX_DISPLAY

studio/frontend/src/features/chat/api/chat-adapter.ts

@@ -2,7 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import type { ChatModelAdapter } from "@assistant-ui/react";
-import type { MessageTiming } from "@assistant-ui/core";
+import type { MessageTiming, ToolCallMessagePart } from "@assistant-ui/core";
 import { toast } from "sonner";
 import {
   generateAudio,
@@ -527,7 +527,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       let reasoningDuration = 0;
       // Tool call content parts — accumulated and yielded cumulatively.
       // result is set directly on the tool-call part when tool_end arrives.
-      const toolCallParts: { type: "tool-call"; toolCallId: string; toolName: string; args: Record<string, unknown>; result?: unknown }[] = [];
+      const toolCallParts: ToolCallMessagePart[] = [];
 
       try {
         const { supportsReasoning, reasoningEnabled } = runtime;
@@ -582,11 +582,13 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           if (toolEvent !== undefined) {
             if (toolEvent.type === "tool_start") {
               const id = (toolEvent.tool_call_id as string) || `${toolEvent.tool_name}_${Date.now()}`;
+              const toolArgs = (toolEvent.arguments ?? {}) as ToolCallMessagePart["args"];
               toolCallParts.push({
                 type: "tool-call" as const,
                 toolCallId: id,
                 toolName: toolEvent.tool_name as string,
-                args: (toolEvent.arguments as Record<string, unknown>) ?? {},
+                argsText: JSON.stringify(toolArgs),
+                args: toolArgs,
               });
             } else if (toolEvent.type === "tool_end") {
               const id = (toolEvent.tool_call_id as string) ||