diff --git a/frontend/src/components/chat/ChatMessage.tsx b/frontend/src/components/chat/ChatMessage.tsx
index ba353f7..98d1bdb 100644
--- a/frontend/src/components/chat/ChatMessage.tsx
+++ b/frontend/src/components/chat/ChatMessage.tsx
@@ -78,6 +78,7 @@ export function ChatMessage({
       }`}
     >
       <div
+        dir={message.role === "user" ? "auto" : undefined}
         className={`max-w-[80%] break-words ${
           message.role === "user"
             ? "p-3 rounded-lg bg-gray-800 dark:bg-gray-700 text-white rounded-br-none whitespace-pre-wrap"
diff --git a/frontend/src/components/chat/MarkdownRenderer.tsx b/frontend/src/components/chat/MarkdownRenderer.tsx
index 06cfb66..fb914f9 100644
--- a/frontend/src/components/chat/MarkdownRenderer.tsx
+++ b/frontend/src/components/chat/MarkdownRenderer.tsx
@@ -8,6 +8,7 @@ import remarkGfm from "remark-gfm";
 import { Prism as SyntaxHighlighter } from "react-syntax-highlighter";
 import { oneLight } from "react-syntax-highlighter/dist/esm/styles/prism";
 import { Copy, Check } from "lucide-react";
+import { detectTextDirection } from "@/lib/utils";
 
 function completePartialMarkdown(text: string): string {
   const fenceCount = (text.match(/^```/gm) || []).length;
@@ -83,7 +84,10 @@ const components: Record<string, any> = {
 export function MarkdownRenderer({ content }: { content: string }) {
   if (!content) return null;
   return (
-    <div className="markdown-body leading-relaxed [&_p]:my-1.5 [&_ul]:my-1.5 [&_ul]:pl-5 [&_ul]:list-disc [&_ol]:my-1.5 [&_ol]:pl-5 [&_ol]:list-decimal [&_li]:my-0.5 [&_h1]:text-lg [&_h1]:font-semibold [&_h1]:mt-3 [&_h1]:mb-1.5 [&_h2]:text-base [&_h2]:font-semibold [&_h2]:mt-2.5 [&_h2]:mb-1 [&_h3]:text-sm [&_h3]:font-semibold [&_h3]:mt-2 [&_h3]:mb-1 [&_blockquote]:border-l-2 [&_blockquote]:border-border [&_blockquote]:pl-3 [&_blockquote]:my-1.5 [&_blockquote]:text-muted-foreground [&_table]:my-2 [&_table]:min-w-full [&_table]:border-collapse [&_table]:text-xs [&_th]:px-2 [&_th]:py-1 [&_th]:bg-muted [&_th]:border [&_th]:border-border [&_th]:text-left [&_th]:font-medium [&_td]:px-2 [&_td]:py-1 [&_td]:border [&_td]:border-border [&>*:first-child]:mt-0 [&>*:last-child]:mb-0">
+    <div
+      dir={detectTextDirection(content)}
+      className="markdown-body leading-relaxed [&_p]:my-1.5 [&_ul]:my-1.5 [&_ul]:ps-5 [&_ul]:list-disc [&_ol]:my-1.5 [&_ol]:ps-5 [&_ol]:list-decimal [&_li]:my-0.5 [&_h1]:text-lg [&_h1]:font-semibold [&_h1]:mt-3 [&_h1]:mb-1.5 [&_h2]:text-base [&_h2]:font-semibold [&_h2]:mt-2.5 [&_h2]:mb-1 [&_h3]:text-sm [&_h3]:font-semibold [&_h3]:mt-2 [&_h3]:mb-1 [&_blockquote]:border-s-2 [&_blockquote]:border-border [&_blockquote]:ps-3 [&_blockquote]:my-1.5 [&_blockquote]:text-muted-foreground [&_table]:my-2 [&_table]:min-w-full [&_table]:border-collapse [&_table]:text-xs [&_th]:px-2 [&_th]:py-1 [&_th]:bg-muted [&_th]:border [&_th]:border-border [&_th]:text-start [&_th]:font-medium [&_td]:px-2 [&_td]:py-1 [&_td]:border [&_td]:border-border [&>*:first-child]:mt-0 [&>*:last-child]:mb-0"
+    >
       <ReactMarkdown remarkPlugins={[remarkGfm]} components={components}>
         {completePartialMarkdown(content)}
       </ReactMarkdown>
diff --git a/frontend/src/components/chat/ReportMarkdownRenderer.tsx b/frontend/src/components/chat/ReportMarkdownRenderer.tsx
index c8f6863..5d934d1 100644
--- a/frontend/src/components/chat/ReportMarkdownRenderer.tsx
+++ b/frontend/src/components/chat/ReportMarkdownRenderer.tsx
@@ -15,6 +15,7 @@ import {
   ChevronRight,
   ExternalLink,
 } from "lucide-react";
+import { detectTextDirection } from "@/lib/utils";
 
 function completePartialMarkdown(text: string): string {
   const fenceCount = (text.match(/^```/gm) || []).length;
@@ -53,12 +54,30 @@ export function extractReportTitle(content: string): string | null {
   return match ? match[1] : null;
 }
 
+// Unicode-aware so non-Latin headings (e.g. Arabic) keep distinct ids; an
+// ASCII-only slug collapses them all to "-", and the duplicate React keys make
+// in-place section updates render as appended blocks.
+function slugify(title: string): string {
+  return (
+    title
+      .toLowerCase()
+      .replace(/[^\p{L}\p{N}]+/gu, "-")
+      .replace(/^-+|-+$/g, "") || "section"
+  );
+}
+
 // Parse markdown into sections based on headings
-function parseIntoSections(content: string): Section[] {
+export function parseIntoSections(content: string): Section[] {
   const lines = content.split("\n");
   const sections: Section[] = [];
   let currentSection: Section | null = null;
   let contentLines: string[] = [];
+  const idCounts = new Map<string, number>();
+  const uniqueId = (base: string) => {
+    const n = idCounts.get(base) ?? 0;
+    idCounts.set(base, n + 1);
+    return n === 0 ? base : `${base}-${n + 1}`;
+  };
 
   for (const line of lines) {
     // Check H2 first (more specific) before H1
@@ -74,9 +93,12 @@ function parseIntoSections(content: string): Section[] {
 
       const level = h1Match ? 1 : 2;
       const title = h1Match ? h1Match[1] : h2Match![1];
-      const id = title.toLowerCase().replace(/[^a-z0-9]+/g, "-");
-
-      currentSection = { id, level, title, content: "" };
+      currentSection = {
+        id: uniqueId(slugify(title)),
+        level,
+        title,
+        content: "",
+      };
       contentLines = [];
     } else if (currentSection) {
       contentLines.push(line);
@@ -106,7 +128,7 @@ const markdownComponents: Record<string, any> = {
   h2: () => null,
   h3({ children }: { children?: React.ReactNode }) {
     return (
-      <h3 className="text-base font-semibold text-foreground mt-4 mb-2 pl-2 border-l-3 border-blue-300 dark:border-blue-600">
+      <h3 className="text-base font-semibold text-foreground mt-4 mb-2 ps-2 border-s-3 border-blue-300 dark:border-blue-600">
         {children}
       </h3>
     );
@@ -118,7 +140,7 @@ const markdownComponents: Record<string, any> = {
     const text = String(children);
     const isAnchor = href?.startsWith("#");
     const isSource = text.startsWith("http") || text.includes("Source");
-    // Citation superscript links like [1], [2] — open source URL or scroll to ref
+    // Citation superscript links like [1], [2] open the source URL or scroll to the ref
     const isCitation = /^\[\d+\]$/.test(text);
     if (isCitation) {
       return (
@@ -159,14 +181,14 @@ const markdownComponents: Record<string, any> = {
   },
   ul({ children }: { children?: React.ReactNode }) {
     return (
-      <ul className="my-2.5 pl-5 list-disc space-y-1.5 text-foreground">
+      <ul className="my-2.5 ps-5 list-disc space-y-1.5 text-foreground">
         {children}
       </ul>
     );
   },
   ol({ children }: { children?: React.ReactNode }) {
     return (
-      <ol className="my-2.5 pl-5 list-decimal space-y-1.5 text-foreground">
+      <ol className="my-2.5 ps-5 list-decimal space-y-1.5 text-foreground">
         {children}
       </ol>
     );
@@ -176,7 +198,7 @@ const markdownComponents: Record<string, any> = {
   },
   blockquote({ children }: { children?: React.ReactNode }) {
     return (
-      <blockquote className="my-3 pl-4 border-l-4 border-blue-300 dark:border-blue-600 bg-blue-50/50 dark:bg-blue-950/30 py-2 pr-3 rounded-r-lg text-muted-foreground italic">
+      <blockquote className="my-3 ps-4 border-s-4 border-blue-300 dark:border-blue-600 bg-blue-50/50 dark:bg-blue-950/30 py-2 pe-3 rounded-e-lg text-muted-foreground italic">
         {children}
       </blockquote>
     );
@@ -193,7 +215,7 @@ const markdownComponents: Record<string, any> = {
   },
   th({ children }: { children?: React.ReactNode }) {
     return (
-      <th className="px-4 py-2.5 text-left text-xs font-semibold text-muted-foreground uppercase tracking-wider">
+      <th className="px-4 py-2.5 text-start text-xs font-semibold text-muted-foreground uppercase tracking-wider">
         {children}
       </th>
     );
@@ -271,14 +293,14 @@ function CollapsibleSection({
       ? {
           wrapper: "mt-6 mb-4 scroll-mt-4",
           button:
-            "w-full flex items-center gap-2 px-3 py-3 rounded-xl bg-gradient-to-r from-blue-100 via-blue-50 to-transparent dark:from-blue-900/40 dark:via-blue-950/20 dark:to-transparent hover:from-blue-200 dark:hover:from-blue-900/60 transition-all text-left group border-b-2 border-blue-200 dark:border-blue-800",
+            "w-full flex items-center gap-2 px-3 py-3 rounded-xl bg-gradient-to-r rtl:bg-gradient-to-l from-blue-100 via-blue-50 to-transparent dark:from-blue-900/40 dark:via-blue-950/20 dark:to-transparent hover:from-blue-200 dark:hover:from-blue-900/60 transition-all text-start group border-b-2 border-blue-200 dark:border-blue-800",
           text: "text-xl font-bold text-foreground group-hover:text-blue-800 dark:group-hover:text-blue-300 transition-colors",
           icon: "w-5 h-5 text-blue-600",
         }
       : {
           wrapper: "mt-5 mb-3 scroll-mt-4",
           button:
-            "w-full flex items-center gap-2 px-3 py-2.5 rounded-lg bg-gradient-to-r from-blue-50 via-blue-50/50 to-transparent dark:from-blue-950/30 dark:via-blue-950/15 dark:to-transparent hover:from-blue-100 dark:hover:from-blue-950/50 transition-all text-left group",
+            "w-full flex items-center gap-2 px-3 py-2.5 rounded-lg bg-gradient-to-r rtl:bg-gradient-to-l from-blue-50 via-blue-50/50 to-transparent dark:from-blue-950/30 dark:via-blue-950/15 dark:to-transparent hover:from-blue-100 dark:hover:from-blue-950/50 transition-all text-start group",
           text: "text-lg font-semibold text-foreground group-hover:text-blue-700 dark:group-hover:text-blue-300 transition-colors",
           icon: "w-5 h-5 text-blue-500",
         };
@@ -295,19 +317,19 @@ function CollapsibleSection({
     >
       <button onClick={onToggle} className={styles.button}>
         {isCollapsed ? (
-          <ChevronRight className={`${styles.icon} shrink-0`} />
+          <ChevronRight className={`${styles.icon} shrink-0 rtl:rotate-180`} />
         ) : (
           <ChevronDown className={`${styles.icon} shrink-0`} />
         )}
         <span className={styles.text}>{section.title}</span>
         {isNew && (
-          <span className="ml-2 px-2 py-0.5 text-xs font-medium bg-green-100 text-green-700 rounded-full">
+          <span className="ms-2 px-2 py-0.5 text-xs font-medium bg-green-100 text-green-700 rounded-full">
             Updated
           </span>
         )}
       </button>
       {!isCollapsed && section.content && (
-        <div className="pl-6 pt-2">
+        <div className="ps-6 pt-2">
           <ReactMarkdown
             remarkPlugins={[remarkGfm]}
             rehypePlugins={[rehypeRaw]}
@@ -340,6 +362,7 @@ export function ReportMarkdownRenderer({
   const prevSectionsRef = useRef<Map<string, string>>(new Map());
 
   const sections = useMemo(() => parseIntoSections(content), [content]);
+  const direction = useMemo(() => detectTextDirection(content), [content]);
 
   // Track changes between versions
   useEffect(() => {
@@ -414,7 +437,7 @@ export function ReportMarkdownRenderer({
   if (!content) return null;
 
   return (
-    <div className="report-markdown">
+    <div className="report-markdown" dir={direction}>
       {/* Table of Contents */}
       {toc.length > 3 && (
         <nav className="mb-6 p-4 bg-muted/50 rounded-xl border border-border">
@@ -425,15 +448,15 @@ export function ReportMarkdownRenderer({
             {toc.map((section) => (
               <li
                 key={section.id}
-                style={{ paddingLeft: `${(section.level - 1) * 12}px` }}
+                style={{ paddingInlineStart: `${(section.level - 1) * 12}px` }}
               >
                 <button
                   onClick={() => scrollToSection(section.id)}
-                  className="text-sm text-muted-foreground hover:text-blue-600 dark:hover:text-blue-400 transition-colors text-left"
+                  className="text-sm text-muted-foreground hover:text-blue-600 dark:hover:text-blue-400 transition-colors text-start"
                 >
                   {section.title}
                   {changedSectionIds.has(section.id) && (
-                    <span className="ml-2 inline-block w-2 h-2 bg-green-500 rounded-full" />
+                    <span className="ms-2 inline-block w-2 h-2 bg-green-500 rounded-full" />
                   )}
                 </button>
               </li>
@@ -481,7 +504,7 @@ export function ReportMarkdownRenderer({
           // Non-collapsible fallback for H2
           return (
             <div key={section.id} id={section.id} className="mb-4 scroll-mt-4">
-              <h2 className="text-lg font-semibold text-foreground mt-5 mb-3 px-3 py-2 rounded-lg bg-gradient-to-r from-blue-50 to-transparent dark:from-blue-950/30 dark:to-transparent">
+              <h2 className="text-lg font-semibold text-foreground mt-5 mb-3 px-3 py-2 rounded-lg bg-gradient-to-r rtl:bg-gradient-to-l from-blue-50 to-transparent dark:from-blue-950/30 dark:to-transparent">
                 {section.title}
               </h2>
               {section.content && (
diff --git a/frontend/src/components/chat/ResearchReportPanel.tsx b/frontend/src/components/chat/ResearchReportPanel.tsx
index d0d1402..9151e64 100644
--- a/frontend/src/components/chat/ResearchReportPanel.tsx
+++ b/frontend/src/components/chat/ResearchReportPanel.tsx
@@ -131,6 +131,7 @@ export function ResearchReportPanel({
             <FileText className="w-5 h-5 text-blue-600" />
           </div>
           <h2
+            dir="auto"
             className="font-semibold text-foreground truncate"
             title={reportTitle}
           >
diff --git a/frontend/src/lib/utils.ts b/frontend/src/lib/utils.ts
index 475ccdf..e03645d 100644
--- a/frontend/src/lib/utils.ts
+++ b/frontend/src/lib/utils.ts
@@ -6,3 +6,15 @@ import { twMerge } from "tailwind-merge";
 export function cn(...inputs: ClassValue[]) {
   return twMerge(clsx(inputs));
 }
+
+const RTL_CHARS = /[\u0590-\u08FF\uFB1D-\uFDFF\uFE70-\uFEFF]/g;
+const LTR_CHARS = /[A-Za-z\u00C0-\u024F]/g;
+
+// Base direction for a block of (markdown) text: RTL when Hebrew/Arabic letters
+// outnumber Latin ones. URLs are ignored so citations don't skew the count.
+export function detectTextDirection(text: string): "rtl" | "ltr" {
+  const stripped = text.replace(/https?:\/\/\S+/g, "");
+  const rtl = stripped.match(RTL_CHARS)?.length ?? 0;
+  const ltr = stripped.match(LTR_CHARS)?.length ?? 0;
+  return rtl > ltr ? "rtl" : "ltr";
+}
diff --git a/frontend/src/test/rtl-report.test.ts b/frontend/src/test/rtl-report.test.ts
new file mode 100644
index 0000000..e6abbac
--- /dev/null
+++ b/frontend/src/test/rtl-report.test.ts
@@ -0,0 +1,68 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT-0
+
+import { describe, it, expect } from "vitest";
+import { parseIntoSections } from "../components/chat/ReportMarkdownRenderer";
+import { detectTextDirection } from "../lib/utils";
+
+const ARABIC_SKELETON = `# مستقبل الذكاء الاصطناعي
+
+## الملخص التنفيذي
+[placeholder]
+
+## الخلفية
+[placeholder]
+
+## أهم النتائج
+
+### النتيجة 1: المسار التقني
+[placeholder]
+
+## التحليل
+[placeholder]`;
+
+describe("Report section parsing", () => {
+  it("gives Arabic headings distinct ids", () => {
+    const ids = parseIntoSections(ARABIC_SKELETON).map((s) => s.id);
+    expect(new Set(ids).size).toBe(ids.length);
+    expect(ids).toContain("الملخص-التنفيذي");
+  });
+
+  it("keeps section ids stable when a placeholder is filled in", () => {
+    const before = parseIntoSections(ARABIC_SKELETON).map((s) => s.id);
+    const filled = ARABIC_SKELETON.replace(
+      "## الخلفية\n[placeholder]",
+      "## الخلفية\nنص جديد للقسم.",
+    );
+    const after = parseIntoSections(filled);
+    expect(after.map((s) => s.id)).toEqual(before);
+    expect(after.find((s) => s.title === "الخلفية")?.content).toBe(
+      "نص جديد للقسم.",
+    );
+  });
+
+  it("disambiguates repeated headings", () => {
+    const ids = parseIntoSections("## Notes\na\n## Notes\nb").map((s) => s.id);
+    expect(ids).toEqual(["notes", "notes-2"]);
+  });
+
+  it("keeps the existing slug for English headings", () => {
+    const [section] = parseIntoSections("## Executive Summary\ntext");
+    expect(section.id).toBe("executive-summary");
+  });
+});
+
+describe("detectTextDirection", () => {
+  it("detects Arabic text with embedded English terms as rtl", () => {
+    expect(
+      detectTextDirection(
+        "تتوقع شركة McKinsey أن يضيف الذكاء الاصطناعي (AGI) قيمة كبيرة [Source: https://www.mckinsey.com/featured-insights]",
+      ),
+    ).toBe("rtl");
+  });
+
+  it("detects English text as ltr", () => {
+    expect(detectTextDirection("The market grew by 15% in 2025.")).toBe("ltr");
+    expect(detectTextDirection("")).toBe("ltr");
+  });
+});
diff --git a/patterns/strands-deep-research/pdf_generator.py b/patterns/strands-deep-research/pdf_generator.py
index 01ee391..ce934e8 100644
--- a/patterns/strands-deep-research/pdf_generator.py
+++ b/patterns/strands-deep-research/pdf_generator.py
@@ -31,7 +31,7 @@
     margin-bottom: 0.4em;
     border-bottom: 2px solid #333;
     padding-bottom: 0.2em;
-    text-align: left;
+    text-align: start;
 }
 h2 {
     font-size: 16pt;
@@ -39,10 +39,10 @@
     margin-bottom: 0.4em;
     border-bottom: 1px solid #ccc;
     padding-bottom: 0.15em;
-    text-align: left;
+    text-align: start;
 }
-h3 { font-size: 13pt; margin-top: 1em; margin-bottom: 0.3em; text-align: left; }
-h4 { font-size: 11pt; margin-top: 0.8em; margin-bottom: 0.3em; text-align: left; }
+h3 { font-size: 13pt; margin-top: 1em; margin-bottom: 0.3em; text-align: start; }
+h4 { font-size: 11pt; margin-top: 0.8em; margin-bottom: 0.3em; text-align: start; }
 p { margin: 0.4em 0; }
 ul, ol { margin: 0.4em 0; padding-left: 1.5em; }
 li { margin: 0.15em 0; }
@@ -55,7 +55,7 @@
 th, td {
     border: 1px solid #ccc;
     padding: 5px 8px;
-    text-align: left;
+    text-align: start;
     font-size: 9pt;
 }
 th { background: #f0f0f0; font-weight: bold; }
@@ -105,6 +105,15 @@
 .references h2 { border-bottom: none; font-size: 14pt; }
 .references ol { font-size: 9pt; padding-left: 1.5em; }
 .references li { margin: 0.3em 0; word-break: break-all; }
+/* Right-to-left documents (Arabic, Hebrew): mirror the box-model offsets */
+html[dir="rtl"] body { direction: rtl; }
+html[dir="rtl"] ul, html[dir="rtl"] ol,
+html[dir="rtl"] .references ol { padding-left: 0; padding-right: 1.5em; }
+html[dir="rtl"] blockquote {
+    border-left: none;
+    border-right: 3px solid #ccc;
+}
+html[dir="rtl"] .references li a { direction: ltr; unicode-bidi: embed; }
 """
 
 _MD_EXTENSIONS = ["tables", "fenced_code", "toc", "nl2br"]
@@ -112,6 +121,17 @@
 # matches [Source: ...] patterns (URL or plain text)
 _SOURCE_RE = re.compile(r"\[Source:\s*([^\]]+)\]")
 _URL_RE = re.compile(r"^https?://\S+$")
+_ANY_URL_RE = re.compile(r"https?://\S+")
+# Hebrew, Arabic (incl. supplements/extended) and their presentation forms
+_RTL_CHARS_RE = re.compile(r"[\u0590-\u08FF\uFB1D-\uFDFF\uFE70-\uFEFF]")
+_LTR_CHARS_RE = re.compile(r"[A-Za-z\u00C0-\u024F]")
+
+
+def _text_direction(text: str) -> str:
+    """Return "rtl" when Hebrew/Arabic letters outnumber Latin ones, else "ltr"."""
+    text = _ANY_URL_RE.sub("", text)
+    rtl = len(_RTL_CHARS_RE.findall(text))
+    return "rtl" if rtl > len(_LTR_CHARS_RE.findall(text)) else "ltr"
 
 
 def _process_citations(html_body: str) -> str:
@@ -162,9 +182,10 @@ def generate_pdf(markdown_content: str) -> bytes:
     """
     html_body = markdown.markdown(markdown_content, extensions=_MD_EXTENSIONS)
     html_body = _process_citations(html_body)
+    direction = _text_direction(markdown_content)
     html_doc = (
         "<!DOCTYPE html>"
-        '<html><head><meta charset="utf-8">'
+        f'<html dir="{direction}"><head><meta charset="utf-8">'
         f"<style>{_CSS}</style></head>"
         f"<body>{html_body}</body></html>"
     )
diff --git a/tools/code_interpreter/execute_python_tool.py b/tools/code_interpreter/execute_python_tool.py
index 9ef9857..a8816b8 100644
--- a/tools/code_interpreter/execute_python_tool.py
+++ b/tools/code_interpreter/execute_python_tool.py
@@ -6,6 +6,7 @@
 import json
 import os
 import re
+from pathlib import Path
 
 import boto3
 from botocore.config import Config
@@ -19,6 +20,13 @@
 STAGING_BUCKET = os.environ.get("STAGING_BUCKET_NAME", "")
 URL_EXPIRATION = 3600
 
+# Run rtl_text.py in the sandbox before the agent's code so matplotlib renders
+# Arabic/Hebrew labels shaped and in the right order. Kept to one line so error
+# line numbers in the agent's code shift by exactly one.
+_RTL_PRELUDE = "exec({src!r}, {{'__name__': '_rtl_text'}})\n".format(
+    src=(Path(__file__).parent / "rtl_text.py").read_text(encoding="utf-8")
+)
+
 
 def _get_interpreter() -> CodeInterpreterTools:
     global _interpreter
@@ -63,6 +71,9 @@ def execute_python(code: str, chart_name: str = "") -> str:
     Do NOT use this tool during the research or report writing phases (Steps 1-4).
 
     The sandbox has numpy, pandas, matplotlib, seaborn, scipy, and scikit-learn.
+    Arabic and Hebrew text in matplotlib titles, labels, ticks and legends is
+    shaped and ordered automatically: pass plain strings, do not reverse them or
+    use arabic_reshaper/bidi (they are not installed).
 
     IMPORTANT: The sandbox is remote. To output a chart, save to BytesIO and print base64:
 
@@ -80,7 +91,7 @@ def execute_python(code: str, chart_name: str = "") -> str:
         code: Python code for data analysis or visualization. Include all imports.
         chart_name: Short filename for the chart (e.g., "gold_forecasts").
     """  # noqa: E501
-    result = _get_interpreter().execute_python_securely(code)
+    result = _get_interpreter().execute_python_securely(_RTL_PRELUDE + code)
 
     if STAGING_BUCKET and chart_name:
         try:
diff --git a/tools/code_interpreter/rtl_text.py b/tools/code_interpreter/rtl_text.py
new file mode 100644
index 0000000..c0cc1a6
--- /dev/null
+++ b/tools/code_interpreter/rtl_text.py
@@ -0,0 +1,308 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: MIT-0
+"""Right-to-left (Arabic/Hebrew) text support for matplotlib.
+
+matplotlib < 3.11 draws glyphs one code point at a time, left to right, with
+no bidi reordering (and, before 3.10, no shaping). Arabic labels therefore come out as disconnected letters in reverse
+order. This module converts a logical string into its visual form (contextual
+Arabic presentation forms, lam-alef ligatures, simplified Unicode bidi
+reordering) and patches ``matplotlib.text.Text.set_text`` to apply it.
+
+The Code Interpreter sandbox has no ``arabic_reshaper``/``python-bidi`` and no
+network access, so this file must stay self-contained (stdlib only). Its source
+is sent to the sandbox and executed ahead of the agent's code; see
+``execute_python_tool.py``.
+"""
+
+import re
+
+# isolated, final, initial, medial. Right-joining letters have no initial/medial.
+_FORMS = {
+    "ء": ("\ufe80", None, None, None),
+    "آ": ("\ufe81", "\ufe82", None, None),
+    "أ": ("\ufe83", "\ufe84", None, None),
+    "ؤ": ("\ufe85", "\ufe86", None, None),
+    "إ": ("\ufe87", "\ufe88", None, None),
+    "ئ": ("\ufe89", "\ufe8a", "\ufe8b", "\ufe8c"),
+    "ا": ("\ufe8d", "\ufe8e", None, None),
+    "ب": ("\ufe8f", "\ufe90", "\ufe91", "\ufe92"),
+    "ة": ("\ufe93", "\ufe94", None, None),
+    "ت": ("\ufe95", "\ufe96", "\ufe97", "\ufe98"),
+    "ث": ("\ufe99", "\ufe9a", "\ufe9b", "\ufe9c"),
+    "ج": ("\ufe9d", "\ufe9e", "\ufe9f", "\ufea0"),
+    "ح": ("\ufea1", "\ufea2", "\ufea3", "\ufea4"),
+    "خ": ("\ufea5", "\ufea6", "\ufea7", "\ufea8"),
+    "د": ("\ufea9", "\ufeaa", None, None),
+    "ذ": ("\ufeab", "\ufeac", None, None),
+    "ر": ("\ufead", "\ufeae", None, None),
+    "ز": ("\ufeaf", "\ufeb0", None, None),
+    "س": ("\ufeb1", "\ufeb2", "\ufeb3", "\ufeb4"),
+    "ش": ("\ufeb5", "\ufeb6", "\ufeb7", "\ufeb8"),
+    "ص": ("\ufeb9", "\ufeba", "\ufebb", "\ufebc"),
+    "ض": ("\ufebd", "\ufebe", "\ufebf", "\ufec0"),
+    "ط": ("\ufec1", "\ufec2", "\ufec3", "\ufec4"),
+    "ظ": ("\ufec5", "\ufec6", "\ufec7", "\ufec8"),
+    "ع": ("\ufec9", "\ufeca", "\ufecb", "\ufecc"),
+    "غ": ("\ufecd", "\ufece", "\ufecf", "\ufed0"),
+    "ف": ("\ufed1", "\ufed2", "\ufed3", "\ufed4"),
+    "ق": ("\ufed5", "\ufed6", "\ufed7", "\ufed8"),
+    "ك": ("\ufed9", "\ufeda", "\ufedb", "\ufedc"),
+    "ل": ("\ufedd", "\ufede", "\ufedf", "\ufee0"),
+    "م": ("\ufee1", "\ufee2", "\ufee3", "\ufee4"),
+    "ن": ("\ufee5", "\ufee6", "\ufee7", "\ufee8"),
+    "ه": ("\ufee9", "\ufeea", "\ufeeb", "\ufeec"),
+    "و": ("\ufeed", "\ufeee", None, None),
+    "ى": ("\ufeef", "\ufef0", None, None),
+    "ي": ("\ufef1", "\ufef2", "\ufef3", "\ufef4"),
+    # Persian/Urdu letters
+    "پ": ("\ufb56", "\ufb57", "\ufb58", "\ufb59"),
+    "چ": ("\ufb7a", "\ufb7b", "\ufb7c", "\ufb7d"),
+    "ژ": ("\ufb8a", "\ufb8b", None, None),
+    "ک": ("\ufb8e", "\ufb8f", "\ufb90", "\ufb91"),
+    "گ": ("\ufb92", "\ufb93", "\ufb94", "\ufb95"),
+    "ی": ("\ufbfc", "\ufbfd", "\ufbfe", "\ufbff"),
+}
+# lam + alef variant -> (isolated, final) ligature
+_LAM_ALEF = {
+    "آ": ("\ufef5", "\ufef6"),
+    "أ": ("\ufef7", "\ufef8"),
+    "إ": ("\ufef9", "\ufefa"),
+    "ا": ("\ufefb", "\ufefc"),
+}
+_LAM = "\u0644"
+_TATWEEL = "\u0640"
+_TRANSPARENT = re.compile(r"[\u0610-\u061A\u064B-\u065F\u0670\u06D6-\u06ED]")
+
+_RTL = re.compile(
+    r"[\u0590-\u065F\u066B-\u06EF\u06FA-\u08FF\uFB1D-\uFDFF\uFE70-\uFEFF]"
+)
+# Arabic letters (bidi class AL) turn following digits into Arabic numbers
+_ARABIC_LETTER = re.compile(
+    r"[\u0600-\u065F\u066B-\u06EF\u06FA-\u08FF\uFB50-\uFDFF\uFE70-\uFEFF]"
+)
+_DIGIT = re.compile(r"[0-9\u0660-\u0669\u06F0-\u06F9]")
+_COMMON_SEPARATORS = set(".,:/")
+_EURO_SEPARATORS = set("+-")
+_NUM_TERMINATORS = set("%#$\u00b0\u066a\u20ac\u00a3")
+_OPEN_BRACKETS = {"(": ")", "[": "]", "{": "}", "\u00ab": "\u00bb"}
+_MIRROR = dict(zip("()[]{}<>\u00ab\u00bb", ")(][}{><\u00bb\u00ab"))
+
+
+def _joins_forward(ch):
+    """True when ``ch`` can connect to the letter that follows it."""
+    return ch == _TATWEEL or (ch in _FORMS and _FORMS[ch][2] is not None)
+
+
+def _joins_backward(ch):
+    return ch == _TATWEEL or (ch in _FORMS and _FORMS[ch][1] is not None)
+
+
+def shape_arabic(text):
+    """Replace Arabic letters with their contextual presentation forms."""
+    chars = list(text)
+    out = []
+    prev = None  # last non-transparent character emitted
+    i = 0
+    while i < len(chars):
+        ch = chars[i]
+        if _TRANSPARENT.match(ch):
+            out.append(ch)
+            i += 1
+            continue
+        joins_prev = prev is not None and _joins_forward(prev)
+        if ch == _LAM and i + 1 < len(chars) and chars[i + 1] in _LAM_ALEF:
+            iso, fin = _LAM_ALEF[chars[i + 1]]
+            out.append(fin if joins_prev else iso)
+            prev = chars[i + 1]  # alef never joins forward
+            i += 2
+            continue
+        if ch not in _FORMS:
+            out.append(ch)
+            prev = ch
+            i += 1
+            continue
+        j = i + 1
+        while j < len(chars) and _TRANSPARENT.match(chars[j]):
+            j += 1
+        nxt = chars[j] if j < len(chars) else None
+        joins_next = _joins_forward(ch) and nxt is not None and _joins_backward(nxt)
+        iso, fin, ini, med = _FORMS[ch]
+        if joins_prev and joins_next:
+            out.append(med)
+        elif joins_prev:
+            out.append(fin or iso)
+        elif joins_next:
+            out.append(ini)
+        else:
+            out.append(iso)
+        prev = ch
+        i += 1
+    return "".join(out)
+
+
+def _resolve_types(line, base):
+    """Bidi classes after the weak- and neutral-type rules (W1-W7, N1-N2).
+
+    Returns one of "R", "L", "EN" (European number) or "AN" (Arabic number)
+    per character, for a paragraph whose direction is ``base`` ("R" or "L").
+    """
+    types = []
+    last_strong = base  # start of line takes the paragraph direction
+    for ch in line:
+        if _RTL.match(ch):
+            last_strong = "AL" if _ARABIC_LETTER.match(ch) else "R"
+            types.append("R")
+        elif _DIGIT.match(ch):
+            # W2: digits after an Arabic letter are Arabic numbers
+            types.append("AN" if last_strong == "AL" else "EN")
+        elif ch.isalpha():
+            last_strong = "L"
+            types.append("L")
+        else:
+            types.append("N")
+    n = len(types)
+    # W4: a single separator between two numbers of the same kind joins them
+    for i in range(1, n - 1):
+        if types[i] != "N" or types[i - 1] != types[i + 1]:
+            continue
+        ch, num = line[i], types[i - 1]
+        if (num == "EN" and ch in _COMMON_SEPARATORS | _EURO_SEPARATORS) or (
+            num == "AN" and ch in _COMMON_SEPARATORS
+        ):
+            types[i] = num
+    # W5: terminators (%, $, ...) next to European numbers belong to them
+    for i in range(n):
+        if types[i] == "EN":
+            for step in (-1, 1):
+                j = i + step
+                while 0 <= j < n and types[j] == "N" and line[j] in _NUM_TERMINATORS:
+                    types[j] = "EN"
+                    j += step
+    # W7: European numbers inside Latin text behave as Latin
+    last_strong = base
+    for i in range(n):
+        if types[i] in ("R", "L"):
+            last_strong = types[i]
+        elif types[i] == "EN" and last_strong == "L":
+            types[i] = "L"
+    _resolve_brackets(line, types, base)
+    # N1/N2: neutrals between same-direction text take it (numbers count as R),
+    # otherwise the paragraph direction
+    i = 0
+    while i < n:
+        if types[i] != "N":
+            i += 1
+            continue
+        j = i
+        while j < n and types[j] == "N":
+            j += 1
+        before = types[i - 1] if i > 0 else base
+        after = types[j] if j < n else base
+        before, after = (("R" if t in ("EN", "AN") else t) for t in (before, after))
+        resolved = before if before == after else base
+        for k in range(i, j):
+            types[k] = resolved
+        i = j
+    return types
+
+
+def _strong(t):
+    return "R" if t in ("R", "EN", "AN") else t
+
+
+def _resolve_brackets(line, types, base):
+    """N0: a bracket pair takes one direction so both halves mirror together."""
+    pairs, stack = [], []
+    for i, ch in enumerate(line):
+        if ch in _OPEN_BRACKETS:
+            stack.append((i, _OPEN_BRACKETS[ch]))
+        elif stack and ch == stack[-1][1]:
+            pairs.append((stack.pop()[0], i))
+    for start, end in sorted(pairs):
+        inside = {_strong(t) for t in types[start + 1 : end]} - {"N"}
+        if base in inside:
+            direction = base
+        elif inside:
+            # only the opposite direction inside: follow the preceding context
+            direction = base
+            for t in reversed(types[:start]):
+                if t != "N":
+                    direction = _strong(t)
+                    break
+            if direction not in inside:
+                direction = base
+        else:
+            continue
+        types[start] = types[end] = direction
+
+
+def visual_order(line):
+    """Reorder one logical line for left-to-right drawing.
+
+    Implements the subset of the Unicode bidi algorithm that chart labels need:
+    Latin words and numbers stay left-to-right inside right-to-left text,
+    neutrals take the direction of their surroundings, brackets are mirrored.
+    The paragraph direction follows the first strong letter (bidi rule P2).
+    """
+    if not _RTL.search(line):
+        return line
+    first = next((c for c in line if c.isalpha()), "")
+    base = "R" if _RTL.match(first) else "L"
+    if base == "R":
+        level_of = {"R": 1, "L": 2, "EN": 2, "AN": 2}
+    else:
+        level_of = {"R": 1, "L": 0, "EN": 2, "AN": 2}
+    levels = [level_of[t] for t in _resolve_types(line, base)]
+    n = len(line)
+    chars = [_MIRROR.get(c, c) if lvl % 2 else c for c, lvl in zip(line, levels)]
+    for lvl in (2, 1):
+        i = 0
+        while i < n:
+            if levels[i] < lvl:
+                i += 1
+                continue
+            j = i
+            while j < n and levels[j] >= lvl:
+                j += 1
+            chars[i:j] = chars[i:j][::-1]
+            i = j
+    return "".join(chars)
+
+
+def to_display(text):
+    """Shape and reorder each line of ``text`` for matplotlib rendering."""
+    return "\n".join(visual_order(shape_arabic(line)) for line in text.split("\n"))
+
+
+def install_matplotlib_patch():
+    """Make every matplotlib Text (titles, labels, ticks, legends) RTL-aware.
+
+    No-op on matplotlib >= 3.11, which shapes and reorders RTL text itself;
+    converting the text there as well would scramble it again.
+    """
+    import matplotlib
+    import matplotlib.text as mtext
+
+    version = tuple(int(x) for x in re.findall(r"\d+", matplotlib.__version__)[:2])
+    if version >= (3, 11) or getattr(mtext.Text.set_text, "_rtl_patched", False):
+        return
+    original = mtext.Text.set_text
+
+    def set_text(self, s):
+        # skip text already converted (e.g. set_text(get_text())) and mathtext
+        if (
+            isinstance(s, str)
+            and s != getattr(self, "_rtl_display", None)
+            and _RTL.search(s)
+            and s.count("$") < 2
+        ):
+            s = to_display(s)
+            self._rtl_display = s
+        return original(self, s)
+
+    set_text._rtl_patched = True
+    mtext.Text.set_text = set_text
+
+
+if __name__ == "_rtl_text":
+    install_matplotlib_patch()