docs: update changelog and version for v1.13.0a5 (#5200)

Bumps the security-updates group with 2 updates in the / directory: [nltk](https://github.com/nltk/nltk) and [pypdf](https://github.com/py-pdf/pypdf).


Updates `nltk` from 3.9.3 to 3.9.4
- [Changelog](https://github.com/nltk/nltk/blob/develop/ChangeLog)
- [Commits](https://github.com/nltk/nltk/compare/3.9.3...3.9.4)

Updates `pypdf` from 6.9.1 to 6.9.2
- [Release notes](https://github.com/py-pdf/pypdf/releases)
- [Changelog](https://github.com/py-pdf/pypdf/blob/main/CHANGELOG.md)
- [Commits](https://github.com/py-pdf/pypdf/compare/6.9.1...6.9.2)

---
updated-dependencies:
- dependency-name: nltk
  dependency-version: 3.9.4
  dependency-type: indirect
  dependency-group: security-updates
- dependency-name: pypdf
  dependency-version: 6.9.2
  dependency-type: indirect
  dependency-group: security-updates
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

.github/workflows/docs-broken-links.yml

@@ -23,7 +23,7 @@ jobs:
       - name: Set up Node
         uses: actions/setup-node@v4
         with:
-          node-version: "latest"
+          node-version: "22"
 
       - name: Install Mintlify CLI
         run: npm i -g mintlify

docs/ar/changelog.mdx

@@ -4,6 +4,127 @@ description: "تحديثات المنتج والتحسينات وإصلاحات
 icon: "clock"
 mode: "wide"
 ---
+<Update label="31 مارس 2026">
+  ## v1.13.0a5
+
+  [عرض الإصدار على GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a5)
+
+  ## ما الذي تغير
+
+  ### الوثائق
+  - تحديث سجل التغييرات والإصدار لـ v1.13.0a4
+
+  ## المساهمون
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="1 أبريل 2026">
+  ## v1.13.0a4
+
+  [عرض الإصدار على GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a4)
+
+  ## ما الذي تغير
+
+  ### الوثائق
+  - تحديث سجل التغييرات والإصدار لـ v1.13.0a3
+
+  ## المساهمون
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="1 أبريل 2026">
+  ## v1.13.0a3
+
+  [عرض الإصدار على GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a3)
+
+  ## ما الذي تغير
+
+  ### الميزات
+  - إصدار بيانات استخدام الرمز في LLMCallCompletedEvent
+  - استخراج ونشر بيانات الأداة إلى AMP
+
+  ### إصلاح الأخطاء
+  - التعامل مع نماذج GPT-5.x التي لا تدعم معلمة API `stop`
+
+  ### الوثائق
+  - إصلاح عدم الدقة في قدرات الوكيل عبر جميع اللغات
+  - إضافة نظرة عامة على قدرات الوكيل وتحسين وثائق المهارات
+  - إضافة دليل شامل لتكوين SSO
+  - تحديث سجل التغييرات والإصدار لـ v1.13.0rc1
+
+  ### إعادة الهيكلة
+  - تحويل Flow إلى Pydantic BaseModel
+  - تحويل فئات LLM إلى Pydantic BaseModel
+  - استبدال InstanceOf[T] بتعليقات نوع عادية
+  - إزالة الطرق غير المستخدمة
+
+  ## المساهمون
+
+  @dependabot[bot], @greysonlalonde, @iris-clawd, @lorenzejay, @lucasgomide, @thiagomoretto
+
+</Update>
+
+<Update label="27 مارس 2026">
+  ## v1.13.0rc1
+
+  [عرض الإصدار على GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0rc1)
+
+  ## ما الذي تغير
+
+  ### الوثائق
+  - تحديث سجل التغييرات والإصدار لـ v1.13.0a2
+
+  ## المساهمون
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="27 مارس 2026">
+  ## v1.13.0a2
+
+  [عرض الإصدار على GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a2)
+
+  ## ما الذي تغير
+
+  ### الميزات
+  - تحديث تلقائي لمستودع اختبار النشر أثناء الإصدار
+  - تحسين مرونة إصدار المؤسسات وتجربة المستخدم
+
+  ### الوثائق
+  - تحديث سجل التغييرات والإصدار للإصدار v1.13.0a1
+
+  ## المساهمون
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="27 مارس 2026">
+  ## v1.13.0a1
+
+  [عرض الإصدار على GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a1)
+
+  ## ما الذي تغير
+
+  ### إصلاحات الأخطاء
+  - إصلاح الروابط المعطلة في سير العمل الوثائقي عن طريق تثبيت Node على LTS 22
+  - مسح ذاكرة التخزين المؤقت لـ uv للحزم المنشورة حديثًا في الإصدار المؤسسي
+
+  ### الوثائق
+  - إضافة مصفوفة شاملة لأذونات RBAC ودليل النشر
+  - تحديث سجل التغييرات والإصدار للإصدار v1.12.2
+
+  ## المساهمون
+
+  @greysonlalonde, @iris-clawd, @joaomdmoura
+
+</Update>
+
 <Update label="25 مارس 2026">
   ## v1.12.2
 

docs/ar/concepts/agent-capabilities.mdx

@@ -0,0 +1,147 @@
+---
+title: "قدرات الوكيل"
+description: "فهم الطرق الخمس لتوسيع وكلاء CrewAI: الأدوات، MCP، التطبيقات، المهارات، والمعرفة."
+icon: puzzle-piece
+mode: "wide"
+---
+
+## نظرة عامة
+
+يمكن توسيع وكلاء CrewAI بـ **خمسة أنواع مميزة من القدرات**، كل منها يخدم غرضًا مختلفًا. فهم متى تستخدم كل نوع — وكيف يعملون معًا — هو المفتاح لبناء وكلاء فعّالين.
+
+<CardGroup cols={2}>
+  <Card title="الأدوات" icon="wrench" href="/ar/concepts/tools" color="#3B82F6">
+    **دوال قابلة للاستدعاء** — تمنح الوكلاء القدرة على اتخاذ إجراءات. البحث على الويب، عمليات الملفات، استدعاءات API، تنفيذ الكود.
+  </Card>
+  <Card title="خوادم MCP" icon="plug" href="/ar/mcp/overview" color="#8B5CF6">
+    **خوادم أدوات عن بُعد** — تربط الوكلاء بخوادم أدوات خارجية عبر Model Context Protocol. نفس تأثير الأدوات، لكن مستضافة خارجيًا.
+  </Card>
+  <Card title="التطبيقات" icon="grid-2" color="#EC4899">
+    **تكاملات المنصة** — تربط الوكلاء بتطبيقات SaaS (Gmail، Slack، Jira، Salesforce) عبر منصة CrewAI. تعمل محليًا مع رمز تكامل المنصة.
+  </Card>
+  <Card title="المهارات" icon="bolt" href="/ar/concepts/skills" color="#F59E0B">
+    **خبرة المجال** — تحقن التعليمات والإرشادات والمواد المرجعية في إرشادات الوكلاء. المهارات تخبر الوكلاء *كيف يفكرون*.
+  </Card>
+  <Card title="المعرفة" icon="book" href="/ar/concepts/knowledge" color="#10B981">
+    **حقائق مُسترجعة** — توفر للوكلاء بيانات من المستندات والملفات وعناوين URL عبر البحث الدلالي (RAG). المعرفة تعطي الوكلاء *ما يحتاجون معرفته*.
+  </Card>
+</CardGroup>
+
+---
+
+## التمييز الأساسي
+
+أهم شيء يجب فهمه: **هذه القدرات تنقسم إلى فئتين**.
+
+### قدرات الإجراء (الأدوات، MCP، التطبيقات)
+
+تمنح الوكلاء القدرة على **فعل أشياء** — استدعاء APIs، قراءة الملفات، البحث على الويب، إرسال رسائل البريد الإلكتروني. عند التنفيذ، تتحول الأنواع الثلاثة إلى نفس التنسيق الداخلي (مثيلات `BaseTool`) وتظهر في قائمة أدوات موحدة يمكن للوكيل استدعاؤها.
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool
+
+agent = Agent(
+    role="Researcher",
+    goal="Find and compile market data",
+    backstory="Expert market analyst",
+    tools=[SerperDevTool(), FileReadTool()],  # أدوات محلية
+    mcps=["https://mcp.example.com/sse"],     # أدوات خادم MCP عن بُعد
+    apps=["gmail", "google_sheets"],           # تكاملات المنصة
+)
+```
+
+### قدرات السياق (المهارات، المعرفة)
+
+تُعدّل **إرشادات** الوكيل — بحقن الخبرة أو التعليمات أو البيانات المُسترجعة قبل أن يبدأ الوكيل في التفكير. لا تمنح الوكلاء إجراءات جديدة؛ بل تُشكّل كيف يفكر الوكلاء وما هي المعلومات التي يمكنهم الوصول إليها.
+
+```python
+from crewai import Agent
+
+agent = Agent(
+    role="Security Auditor",
+    goal="Audit cloud infrastructure for vulnerabilities",
+    backstory="Expert in cloud security with 10 years of experience",
+    skills=["./skills/security-audit"],        # تعليمات المجال
+    knowledge_sources=[pdf_source, url_source], # حقائق مُسترجعة
+)
+```
+
+---
+
+## متى تستخدم ماذا
+
+| تحتاج إلى...                                            | استخدم            | مثال                                     |
+| :------------------------------------------------------- | :---------------- | :--------------------------------------- |
+| الوكيل يبحث على الويب                                    | **الأدوات**       | `tools=[SerperDevTool()]`                |
+| الوكيل يستدعي API عن بُعد عبر MCP                        | **MCP**           | `mcps=["https://api.example.com/sse"]`   |
+| الوكيل يرسل بريد إلكتروني عبر Gmail                      | **التطبيقات**     | `apps=["gmail"]`                         |
+| الوكيل يتبع إجراءات محددة                                | **المهارات**      | `skills=["./skills/code-review"]`        |
+| الوكيل يرجع لمستندات الشركة                              | **المعرفة**       | `knowledge_sources=[pdf_source]`         |
+| الوكيل يبحث على الويب ويتبع إرشادات المراجعة              | **الأدوات + المهارات** | استخدم كليهما معًا                  |
+
+---
+
+## دمج القدرات
+
+في الممارسة العملية، غالبًا ما يستخدم الوكلاء **أنواعًا متعددة من القدرات معًا**. إليك مثال واقعي:
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+# وكيل بحث مجهز بالكامل
+researcher = Agent(
+    role="Senior Research Analyst",
+    goal="Produce comprehensive market analysis reports",
+    backstory="Expert analyst with deep industry knowledge",
+
+    # الإجراء: ما يمكن للوكيل فعله
+    tools=[
+        SerperDevTool(),         # البحث على الويب
+        FileReadTool(),          # قراءة الملفات المحلية
+        CodeInterpreterTool(),   # تشغيل كود Python للتحليل
+    ],
+    mcps=["https://data-api.example.com/sse"],  # الوصول لـ API بيانات عن بُعد
+    apps=["google_sheets"],                      # الكتابة في Google Sheets
+
+    # السياق: ما يعرفه الوكيل
+    skills=["./skills/research-methodology"],    # كيفية إجراء البحث
+    knowledge_sources=[company_docs],            # بيانات خاصة بالشركة
+)
+```
+
+---
+
+## جدول المقارنة
+
+| الميزة | الأدوات | MCP | التطبيقات | المهارات | المعرفة |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **يمنح الوكيل إجراءات** | ✅ | ✅ | ✅ | ❌ | ❌ |
+| **يُعدّل الإرشادات** | ❌ | ❌ | ❌ | ✅ | ✅ |
+| **يتطلب كود** | نعم | إعداد فقط | إعداد فقط | Markdown فقط | إعداد فقط |
+| **يعمل محليًا** | نعم | يعتمد | نعم (مع متغير بيئة) | غير متاح | نعم |
+| **يحتاج مفاتيح API** | لكل أداة | لكل خادم | رمز التكامل | لا | المُضمّن فقط |
+| **يُعيَّن على Agent** | `tools=[]` | `mcps=[]` | `apps=[]` | `skills=[]` | `knowledge_sources=[]` |
+| **يُعيَّن على Crew** | ❌ | ❌ | ❌ | `skills=[]` | `knowledge_sources=[]` |
+
+---
+
+## تعمّق أكثر
+
+هل أنت مستعد لمعرفة المزيد عن كل نوع من أنواع القدرات؟
+
+<CardGroup cols={2}>
+  <Card title="الأدوات" icon="wrench" href="/ar/concepts/tools">
+    إنشاء أدوات مخصصة، استخدام كتالوج OSS مع أكثر من 75 خيارًا، تكوين التخزين المؤقت والتنفيذ غير المتزامن.
+  </Card>
+  <Card title="تكامل MCP" icon="plug" href="/ar/mcp/overview">
+    الاتصال بخوادم MCP عبر stdio أو SSE أو HTTP. تصفية الأدوات، تكوين المصادقة.
+  </Card>
+  <Card title="المهارات" icon="bolt" href="/ar/concepts/skills">
+    بناء حزم المهارات مع SKILL.md، حقن خبرة المجال، استخدام الكشف التدريجي.
+  </Card>
+  <Card title="المعرفة" icon="book" href="/ar/concepts/knowledge">
+    إضافة المعرفة من ملفات PDF وCSV وعناوين URL والمزيد. تكوين المُضمّنات والاسترجاع.
+  </Card>
+</CardGroup>

docs/ar/concepts/skills.mdx

@@ -1,15 +1,217 @@
 ---
 title: المهارات
-description: حزم المهارات المبنية على نظام الملفات التي تحقن السياق في إرشادات الوكيل.
+description: حزم المهارات المبنية على نظام الملفات التي تحقن خبرة المجال والتعليمات في إرشادات الوكلاء.
 icon: bolt
 mode: "wide"
 ---
 
 ## نظرة عامة
 
-المهارات هي مجلدات مستقلة توفر للوكلاء تعليمات ومراجع وموارد خاصة بالمجال. تُعرّف كل مهارة بملف `SKILL.md` يحتوي على بيانات وصفية YAML ومحتوى Markdown.
+المهارات هي مجلدات مستقلة توفر للوكلاء **تعليمات وإرشادات ومواد مرجعية خاصة بالمجال**. تُعرّف كل مهارة بملف `SKILL.md` يحتوي على بيانات وصفية YAML ومحتوى Markdown.
 
-تستخدم المهارات **الكشف التدريجي** — يتم تحميل البيانات الوصفية أولاً، ثم التعليمات الكاملة فقط عند التفعيل، وكتالوجات الموارد فقط عند الحاجة.
+عند التفعيل، يتم حقن تعليمات المهارة مباشرة في إرشادات مهمة الوكيل — مما يمنح الوكيل خبرة دون الحاجة لأي تغييرات في الكود.
+
+<Note type="info" title="المهارات مقابل الأدوات — التمييز الأساسي">
+**المهارات ليست أدوات.** هذه هي نقطة الارتباك الأكثر شيوعًا.
+
+- **المهارات** تحقن *تعليمات وسياق* في إرشادات الوكيل. تخبر الوكيل *كيف يفكر* في مشكلة ما.
+- **الأدوات** تمنح الوكيل *دوال قابلة للاستدعاء* لاتخاذ إجراءات (البحث، قراءة الملفات، استدعاء APIs).
+
+غالبًا ما تحتاج **كليهما**: مهارات للخبرة، وأدوات للإجراء. يتم تكوينهما بشكل مستقل ويُكمّلان بعضهما.
+</Note>
+
+---
+
+## البداية السريعة
+
+### 1. إنشاء مجلد المهارة
+
+```
+skills/
+└── code-review/
+    ├── SKILL.md          # مطلوب — التعليمات
+    ├── references/       # اختياري — مستندات مرجعية
+    │   └── style-guide.md
+    └── scripts/          # اختياري — سكربتات قابلة للتنفيذ
+```
+
+### 2. كتابة SKILL.md الخاص بك
+
+```markdown
+---
+name: code-review
+description: Guidelines for conducting thorough code reviews with focus on security and performance.
+metadata:
+  author: your-team
+  version: "1.0"
+---
+
+## إرشادات مراجعة الكود
+
+عند مراجعة الكود، اتبع قائمة التحقق هذه:
+
+1. **الأمان**: تحقق من ثغرات الحقن وتجاوز المصادقة وكشف البيانات
+2. **الأداء**: ابحث عن استعلامات N+1 والتخصيصات غير الضرورية والاستدعاءات المحظورة
+3. **القابلية للقراءة**: تأكد من وضوح التسمية والتعليقات المناسبة والأسلوب المتسق
+4. **الاختبارات**: تحقق من تغطية اختبار كافية للوظائف الجديدة
+
+### مستويات الخطورة
+- **حرج**: ثغرات أمنية، مخاطر فقدان البيانات → حظر الدمج
+- **رئيسي**: مشاكل أداء، أخطاء منطقية → طلب تغييرات
+- **ثانوي**: مسائل أسلوبية، اقتراحات تسمية → الموافقة مع تعليقات
+```
+
+### 3. ربطها بوكيل
+
+```python
+from crewai import Agent
+from crewai_tools import GithubSearchTool, FileReadTool
+
+reviewer = Agent(
+    role="Senior Code Reviewer",
+    goal="Review pull requests for quality and security issues",
+    backstory="Staff engineer with expertise in secure coding practices.",
+    skills=["./skills"],                          # يحقن إرشادات المراجعة
+    tools=[GithubSearchTool(), FileReadTool()],   # يسمح للوكيل بقراءة الكود
+)
+```
+
+الوكيل الآن لديه **خبرة** (من المهارة) و**قدرات** (من الأدوات) معًا.
+
+---
+
+## المهارات + الأدوات: العمل معًا
+
+إليك أنماط شائعة توضح كيف تُكمّل المهارات والأدوات بعضهما:
+
+### النمط 1: مهارات فقط (خبرة المجال، بدون إجراءات مطلوبة)
+
+استخدم عندما يحتاج الوكيل لتعليمات محددة لكن لا يحتاج لاستدعاء خدمات خارجية:
+
+```python
+agent = Agent(
+    role="Technical Writer",
+    goal="Write clear API documentation",
+    backstory="Expert technical writer",
+    skills=["./skills/api-docs-style"],  # إرشادات وقوالب الكتابة
+    # لا حاجة لأدوات — الوكيل يكتب بناءً على السياق المقدم
+)
+```
+
+### النمط 2: أدوات فقط (إجراءات، بدون خبرة خاصة)
+
+استخدم عندما يحتاج الوكيل لاتخاذ إجراءات لكن لا يحتاج لتعليمات مجال محددة:
+
+```python
+from crewai_tools import SerperDevTool, ScrapeWebsiteTool
+
+agent = Agent(
+    role="Web Researcher",
+    goal="Find information about a topic",
+    backstory="Skilled at finding information online",
+    tools=[SerperDevTool(), ScrapeWebsiteTool()],  # يمكنه البحث والاستخراج
+    # لا حاجة لمهارات — البحث العام لا يحتاج إرشادات خاصة
+)
+```
+
+### النمط 3: مهارات + أدوات (خبرة وإجراءات)
+
+النمط الأكثر شيوعًا في العالم الحقيقي. المهارة توفر *كيف* تقترب من العمل؛ الأدوات توفر *ما* يمكن للوكيل فعله:
+
+```python
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+analyst = Agent(
+    role="Security Analyst",
+    goal="Audit infrastructure for vulnerabilities",
+    backstory="Expert in cloud security and compliance",
+    skills=["./skills/security-audit"],   # منهجية وقوائم تحقق التدقيق
+    tools=[
+        SerperDevTool(),                  # البحث عن ثغرات معروفة
+        FileReadTool(),                   # قراءة ملفات التكوين
+        CodeInterpreterTool(),            # تشغيل سكربتات التحليل
+    ],
+)
+```
+
+### النمط 4: مهارات + MCP
+
+المهارات تعمل مع خوادم MCP بنفس الطريقة التي تعمل بها مع الأدوات:
+
+```python
+agent = Agent(
+    role="Data Analyst",
+    goal="Analyze customer data and generate reports",
+    backstory="Expert data analyst with strong statistical background",
+    skills=["./skills/data-analysis"],           # منهجية التحليل
+    mcps=["https://data-warehouse.example.com/sse"],  # وصول بيانات عن بُعد
+)
+```
+
+### النمط 5: مهارات + تطبيقات
+
+المهارات يمكن أن توجّه كيف يستخدم الوكيل تكاملات المنصة:
+
+```python
+agent = Agent(
+    role="Customer Support Agent",
+    goal="Respond to customer inquiries professionally",
+    backstory="Experienced support representative",
+    skills=["./skills/support-playbook"],  # قوالب الردود وقواعد التصعيد
+    apps=["gmail", "zendesk"],             # يمكنه إرسال رسائل بريد وتحديث التذاكر
+)
+```
+
+---
+
+## المهارات على مستوى الطاقم
+
+يمكن تعيين المهارات على الطاقم لتُطبّق على **جميع الوكلاء**:
+
+```python
+from crewai import Crew
+
+crew = Crew(
+    agents=[researcher, writer, reviewer],
+    tasks=[research_task, write_task, review_task],
+    skills=["./skills"],  # جميع الوكلاء يحصلون على هذه المهارات
+)
+```
+
+المهارات على مستوى الوكيل لها الأولوية — إذا تم اكتشاف نفس المهارة في كلا المستويين، يتم استخدام نسخة الوكيل.
+
+---
+
+## تنسيق SKILL.md
+
+```markdown
+---
+name: my-skill
+description: وصف قصير لما تفعله هذه المهارة ومتى تُستخدم.
+license: Apache-2.0                    # اختياري
+compatibility: crewai>=0.1.0          # اختياري
+metadata:                              # اختياري
+  author: your-name
+  version: "1.0"
+allowed-tools: web-search file-read   # اختياري، تجريبي
+---
+
+التعليمات للوكيل تُكتب هنا. يتم حقن محتوى Markdown هذا
+في إرشادات الوكيل عند تفعيل المهارة.
+```
+
+### حقول البيانات الوصفية
+
+| الحقل           | مطلوب    | الوصف                                                                    |
+| :-------------- | :------- | :----------------------------------------------------------------------- |
+| `name`          | نعم      | 1-64 حرف. أحرف صغيرة أبجدية رقمية وشرطات. يجب أن يطابق اسم المجلد.     |
+| `description`   | نعم      | 1-1024 حرف. يصف ما تفعله المهارة ومتى تُستخدم.                          |
+| `license`       | لا       | اسم الترخيص أو مرجع لملف ترخيص مضمّن.                                   |
+| `compatibility` | لا       | حد أقصى 500 حرف. متطلبات البيئة (منتجات، حزم، شبكة).                    |
+| `metadata`      | لا       | تعيين مفتاح-قيمة نصي عشوائي.                                            |
+| `allowed-tools` | لا       | قائمة أدوات معتمدة مسبقًا مفصولة بمسافات. تجريبي.                        |
+
+---
 
 ## هيكل المجلد
 
@@ -21,79 +223,25 @@ my-skill/
 └── assets/             # اختياري — ملفات ثابتة (إعدادات، بيانات)
 ```
 
-يجب أن يتطابق اسم المجلد مع حقل `name` في `SKILL.md`.
+يجب أن يتطابق اسم المجلد مع حقل `name` في `SKILL.md`. مجلدات `scripts/` و `references/` و `assets/` متاحة في مسار المهارة `path` للوكلاء الذين يحتاجون للإشارة إلى الملفات مباشرة.
 
-## تنسيق SKILL.md
-
-```markdown
----
-name: my-skill
-description: Short description of what this skill does and when to use it.
-license: Apache-2.0                    # optional
-compatibility: crewai>=0.1.0          # optional
-metadata:                              # optional
-  author: your-name
-  version: "1.0"
-allowed-tools: web-search file-read   # optional, space-delimited
 ---
 
-Instructions for the agent go here. This markdown body is injected
-into the agent's prompt when the skill is activated.
-```
+## المهارات المحمّلة مسبقًا
 
-### حقول البيانات الوصفية
-
-| الحقل           | مطلوب | القيود                                                              |
-| :-------------- | :------- | :----------------------------------------------------------------------- |
-| `name`          | نعم      | 1-64 حرف. أحرف صغيرة أبجدية رقمية وشرطات. بدون شرطات بادئة/لاحقة/متتالية. يجب أن يطابق اسم المجلد. |
-| `description`   | نعم      | 1-1024 حرف. يصف ما تفعله المهارة ومتى تُستخدم.          |
-| `license`       | لا       | اسم الترخيص أو مرجع لملف ترخيص مضمّن.                     |
-| `compatibility` | لا       | حد أقصى 500 حرف. متطلبات البيئة (منتجات، حزم، شبكة).   |
-| `metadata`      | لا       | تعيين مفتاح-قيمة نصي عشوائي.                                      |
-| `allowed-tools` | لا       | قائمة أدوات معتمدة مسبقًا مفصولة بمسافات. تجريبي.                |
-
-## الاستخدام
-
-### المهارات على مستوى الوكيل
-
-مرر مسارات مجلدات المهارات إلى وكيل:
-
-```python
-from crewai import Agent
-
-agent = Agent(
-    role="Researcher",
-    goal="Find relevant information",
-    backstory="An expert researcher.",
-    skills=["./skills"],  # يكتشف جميع المهارات في هذا المجلد
-)
-```
-
-### المهارات على مستوى الطاقم
-
-تُدمج مسارات المهارات في الطاقم مع كل وكيل:
-
-```python
-from crewai import Crew
-
-crew = Crew(
-    agents=[agent],
-    tasks=[task],
-    skills=["./skills"],
-)
-```
-
-### المهارات المحمّلة مسبقًا
-
-يمكنك أيضًا تمرير كائنات `Skill` مباشرة:
+للمزيد من التحكم، يمكنك اكتشاف المهارات وتفعيلها برمجيًا:
 
 ```python
 from pathlib import Path
 from crewai.skills import discover_skills, activate_skill
 
+# اكتشاف جميع المهارات في مجلد
 skills = discover_skills(Path("./skills"))
+
+# تفعيلها (تحميل محتوى SKILL.md الكامل)
 activated = [activate_skill(s) for s in skills]
 
+# تمرير إلى وكيل
 agent = Agent(
     role="Researcher",
     goal="Find relevant information",
@@ -102,13 +250,57 @@ agent = Agent(
 )
 ```
 
+---
+
 ## كيف يتم تحميل المهارات
 
-يتم تحميل المهارات تدريجيًا — فقط البيانات المطلوبة في كل مرحلة يتم قراءتها:
+تستخدم المهارات **الكشف التدريجي** — تحمّل فقط ما هو مطلوب في كل مرحلة:
 
-| المرحلة            | ما يتم تحميله                                     | متى               |
-| :--------------- | :------------------------------------------------ | :----------------- |
-| الاكتشاف        | الاسم، الوصف، حقول البيانات الوصفية             | `discover_skills()` |
-| التفعيل       | نص محتوى SKILL.md الكامل                           | `activate_skill()` |
+| المرحلة    | ما يتم تحميله                          | متى                |
+| :--------- | :------------------------------------ | :------------------ |
+| الاكتشاف  | الاسم، الوصف، حقول البيانات الوصفية   | `discover_skills()` |
+| التفعيل   | نص محتوى SKILL.md الكامل              | `activate_skill()`  |
 
-أثناء التنفيذ العادي للوكيل، يتم اكتشاف المهارات وتفعيلها تلقائيًا. مجلدات `scripts/` و `references/` و `assets/` متاحة في مسار المهارة `path` للوكلاء الذين يحتاجون للإشارة إلى الملفات مباشرة.
+أثناء التنفيذ العادي للوكيل (تمرير مسارات المجلدات عبر `skills=["./skills"]`)، يتم اكتشاف المهارات وتفعيلها تلقائيًا. التحميل التدريجي مهم فقط عند استخدام الواجهة البرمجية.
+
+---
+
+## المهارات مقابل المعرفة
+
+كلا المهارات والمعرفة تُعدّل إرشادات الوكيل، لكنهما يخدمان أغراضًا مختلفة:
+
+| الجانب | المهارات | المعرفة |
+| :--- | :--- | :--- |
+| **ما توفره** | تعليمات، إجراءات، إرشادات | حقائق، بيانات، معلومات |
+| **كيف تُخزّن** | ملفات Markdown (SKILL.md) | مُضمّنة في مخزن متجهي (ChromaDB) |
+| **كيف تُسترجع** | يتم حقن المحتوى الكامل في الإرشادات | البحث الدلالي يجد الأجزاء ذات الصلة |
+| **الأفضل لـ** | المنهجيات، قوائم التحقق، أدلة الأسلوب | مستندات الشركة، معلومات المنتج، بيانات مرجعية |
+| **يُعيّن عبر** | `skills=["./skills"]` | `knowledge_sources=[source]` |
+
+**القاعدة العامة:** إذا كان الوكيل يحتاج لاتباع *عملية*، استخدم مهارة. إذا كان يحتاج للرجوع إلى *بيانات*، استخدم المعرفة.
+
+---
+
+## الأسئلة الشائعة
+
+<AccordionGroup>
+  <Accordion title="هل أحتاج لتعيين المهارات والأدوات معًا؟">
+    يعتمد على حالة الاستخدام. المهارات والأدوات **مستقلتان** — يمكنك استخدام أيّ منهما أو كليهما أو لا شيء.
+
+    - **مهارات فقط**: عندما يحتاج الوكيل خبرة لكن لا يحتاج إجراءات خارجية (مثال: الكتابة بإرشادات أسلوبية)
+    - **أدوات فقط**: عندما يحتاج الوكيل إجراءات لكن لا يحتاج منهجية خاصة (مثال: بحث بسيط على الويب)
+    - **كليهما**: عندما يحتاج الوكيل خبرة وإجراءات (مثال: تدقيق أمني بقوائم تحقق محددة وقدرة على فحص الكود)
+  </Accordion>
+
+  <Accordion title="هل توفر المهارات أدوات تلقائيًا؟">
+    **لا.** حقل `allowed-tools` في SKILL.md هو بيانات وصفية تجريبية فقط — لا يُنشئ أو يحقن أي أدوات. يجب عليك دائمًا تعيين الأدوات بشكل منفصل عبر `tools=[]` أو `mcps=[]` أو `apps=[]`.
+  </Accordion>
+
+  <Accordion title="ماذا يحدث إذا عيّنت نفس المهارة على كل من الوكيل والطاقم؟">
+    المهارة على مستوى الوكيل لها الأولوية. يتم إزالة التكرار حسب الاسم — مهارات الوكيل تُعالج أولاً، لذا إذا ظهر نفس اسم المهارة في كلا المستويين، تُستخدم نسخة الوكيل.
+  </Accordion>
+
+  <Accordion title="ما الحجم الأقصى لمحتوى SKILL.md؟">
+    هناك تحذير ناعم عند 50,000 حرف، لكن بدون حد صارم. حافظ على تركيز المهارات وإيجازها للحصول على أفضل النتائج — الحقن الكبيرة في الإرشادات قد تُشتت انتباه الوكيل.
+  </Accordion>
+</AccordionGroup>

docs/ar/concepts/tools.mdx

@@ -10,6 +10,10 @@ mode: "wide"
 تُمكّن أدوات CrewAI الوكلاء بقدرات تتراوح من البحث على الويب وتحليل البيانات إلى التعاون وتفويض المهام بين الزملاء.
 توضح هذه الوثائق كيفية إنشاء هذه الأدوات ودمجها والاستفادة منها ضمن إطار عمل CrewAI، بما في ذلك التركيز على أدوات التعاون.
 
+<Note type="info" title="الأدوات هي أحد أنواع قدرات الوكيل الخمسة">
+  الأدوات تمنح الوكلاء **دوال قابلة للاستدعاء** لاتخاذ إجراءات. تعمل جنبًا إلى جنب مع [MCP](/ar/mcp/overview) (خوادم أدوات عن بُعد) و[التطبيقات](/ar/concepts/agent-capabilities) (تكاملات المنصة) و[المهارات](/ar/concepts/skills) (خبرة المجال) و[المعرفة](/ar/concepts/knowledge) (حقائق مُسترجعة). راجع نظرة عامة على [قدرات الوكيل](/ar/concepts/agent-capabilities) لفهم متى تستخدم كل نوع.
+</Note>
+
 ## ما هي الأداة؟
 
 الأداة في CrewAI هي مهارة أو وظيفة يمكن للوكلاء استخدامها لأداء إجراءات مختلفة.

docs/ar/mcp/security.mdx

@@ -139,19 +139,7 @@ mode: "wide"
 - **الالتزام بمواصفات ترخيص MCP**: إذا كنت تنفذ المصادقة والترخيص، اتبع بدقة [مواصفات ترخيص MCP](https://modelcontextprotocol.io/specification/draft/basic/authorization).
 - **تدقيقات أمنية منتظمة**: إذا كان خادم MCP يتعامل مع بيانات حساسة، فكر في إجراء تدقيقات أمنية دورية.
 
-## 5. الإبلاغ عن الثغرات الأمنية
-
-إذا اكتشفت ثغرة أمنية في CrewAI، يرجى الإبلاغ عنها بشكل مسؤول من خلال برنامج الكشف عن الثغرات (VDP) الخاص بنا على Bugcrowd:
-
-**أرسل التقارير إلى:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-**لا تكشف** عن الثغرات عبر issues العامة على GitHub أو pull requests أو وسائل التواصل الاجتماعي. لن تتم مراجعة التقارير المقدمة عبر قنوات غير Bugcrowd.
-</Warning>
-
-لمزيد من التفاصيل، راجع [سياسة الأمان](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md) الخاصة بنا.
-
-## 6. قراءة إضافية
+## 5. قراءة إضافية
 
 لمزيد من المعلومات التفصيلية حول أمان MCP، راجع التوثيق الرسمي:
 - **[أمان نقل MCP](https://modelcontextprotocol.io/docs/concepts/transports#security-considerations)**

docs/ar/security.mdx

@@ -1,22 +0,0 @@
----
-title: سياسة الأمان
-description: تعرف على كيفية الإبلاغ عن الثغرات الأمنية وممارسات الأمان في CrewAI.
-icon: shield
-mode: "wide"
----
-
-## الإبلاغ عن الثغرات الأمنية
-
-إذا اكتشفت ثغرة أمنية في CrewAI، يرجى الإبلاغ عنها بشكل مسؤول من خلال برنامج الكشف عن الثغرات (VDP) الخاص بنا على Bugcrowd:
-
-**أرسل التقارير إلى:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-**لا تكشف** عن الثغرات عبر issues العامة على GitHub أو pull requests أو وسائل التواصل الاجتماعي. لن تتم مراجعة التقارير المقدمة عبر قنوات غير Bugcrowd.
-</Warning>
-
-لمزيد من التفاصيل، راجع [سياسة الأمان على GitHub](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md).
-
-## موارد الأمان
-
-- **[اعتبارات أمان MCP](/mcp/security)** — أفضل الممارسات لدمج خوادم MCP بأمان مع وكلاء CrewAI، بما في ذلك أمان النقل ومخاطر حقن الأوامر ونصائح تنفيذ الخادم.

docs/en/changelog.mdx

@@ -4,6 +4,127 @@ description: "Product updates, improvements, and bug fixes for CrewAI"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="Mar 31, 2026">
+  ## v1.13.0a5
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a5)
+
+  ## What's Changed
+
+  ### Documentation
+  - Update changelog and version for v1.13.0a4
+
+  ## Contributors
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="Apr 01, 2026">
+  ## v1.13.0a4
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a4)
+
+  ## What's Changed
+
+  ### Documentation
+  - Update changelog and version for v1.13.0a3
+
+  ## Contributors
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="Apr 01, 2026">
+  ## v1.13.0a3
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a3)
+
+  ## What's Changed
+
+  ### Features
+  - Emit token usage data in LLMCallCompletedEvent
+  - Extract and publish tool metadata to AMP
+
+  ### Bug Fixes
+  - Handle GPT-5.x models not supporting the `stop` API parameter
+
+  ### Documentation
+  - Fix inaccuracies in agent-capabilities across all languages
+  - Add Agent Capabilities overview and improve Skills documentation
+  - Add comprehensive SSO configuration guide
+  - Update changelog and version for v1.13.0rc1
+
+  ### Refactoring
+  - Convert Flow to Pydantic BaseModel
+  - Convert LLM classes to Pydantic BaseModel
+  - Replace InstanceOf[T] with plain type annotations
+  - Remove unused methods
+
+  ## Contributors
+
+  @dependabot[bot], @greysonlalonde, @iris-clawd, @lorenzejay, @lucasgomide, @thiagomoretto
+
+</Update>
+
+<Update label="Mar 27, 2026">
+  ## v1.13.0rc1
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0rc1)
+
+  ## What's Changed
+
+  ### Documentation
+  - Update changelog and version for v1.13.0a2
+
+  ## Contributors
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="Mar 27, 2026">
+  ## v1.13.0a2
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a2)
+
+  ## What's Changed
+
+  ### Features
+  - Auto-update deployment test repo during release
+  - Improve enterprise release resilience and UX
+
+  ### Documentation
+  - Update changelog and version for v1.13.0a1
+
+  ## Contributors
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="Mar 27, 2026">
+  ## v1.13.0a1
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a1)
+
+  ## What's Changed
+
+  ### Bug Fixes
+  - Fix broken links in documentation workflow by pinning Node to LTS 22
+  - Bust the uv cache for freshly published packages in enterprise release
+
+  ### Documentation
+  - Add comprehensive RBAC permissions matrix and deployment guide
+  - Update changelog and version for v1.12.2
+
+  ## Contributors
+
+  @greysonlalonde, @iris-clawd, @joaomdmoura
+
+</Update>
+
 <Update label="Mar 25, 2026">
   ## v1.12.2
 

docs/en/concepts/agent-capabilities.mdx

@@ -0,0 +1,147 @@
+---
+title: "Agent Capabilities"
+description: "Understand the five ways to extend CrewAI agents: Tools, MCPs, Apps, Skills, and Knowledge."
+icon: puzzle-piece
+mode: "wide"
+---
+
+## Overview
+
+CrewAI agents can be extended with **five distinct capability types**, each serving a different purpose. Understanding when to use each one — and how they work together — is key to building effective agents.
+
+<CardGroup cols={2}>
+  <Card title="Tools" icon="wrench" href="/en/concepts/tools" color="#3B82F6">
+    **Callable functions** — give agents the ability to take action. Web searches, file operations, API calls, code execution.
+  </Card>
+  <Card title="MCP Servers" icon="plug" href="/en/mcp/overview" color="#8B5CF6">
+    **Remote tool servers** — connect agents to external tool servers via the Model Context Protocol. Same effect as tools, but hosted externally.
+  </Card>
+  <Card title="Apps" icon="grid-2" color="#EC4899">
+    **Platform integrations** — connect agents to SaaS apps (Gmail, Slack, Jira, Salesforce) via CrewAI's platform. Runs locally with a platform integration token.
+  </Card>
+  <Card title="Skills" icon="bolt" href="/en/concepts/skills" color="#F59E0B">
+    **Domain expertise** — inject instructions, guidelines, and reference material into agent prompts. Skills tell agents *how to think*.
+  </Card>
+  <Card title="Knowledge" icon="book" href="/en/concepts/knowledge" color="#10B981">
+    **Retrieved facts** — provide agents with data from documents, files, and URLs via semantic search (RAG). Knowledge gives agents *what to know*.
+  </Card>
+</CardGroup>
+
+---
+
+## The Key Distinction
+
+The most important thing to understand: **these capabilities fall into two categories**.
+
+### Action Capabilities (Tools, MCPs, Apps)
+
+These give agents the ability to **do things** — call APIs, read files, search the web, send emails. At execution time, all three resolve into the same internal format (`BaseTool` instances) and appear in a unified tool list the agent can call.
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool
+
+agent = Agent(
+    role="Researcher",
+    goal="Find and compile market data",
+    backstory="Expert market analyst",
+    tools=[SerperDevTool(), FileReadTool()],  # Local tools
+    mcps=["https://mcp.example.com/sse"],     # Remote MCP server tools
+    apps=["gmail", "google_sheets"],           # Platform integrations
+)
+```
+
+### Context Capabilities (Skills, Knowledge)
+
+These modify the agent's **prompt** — injecting expertise, instructions, or retrieved data before the agent starts reasoning. They don't give agents new actions; they shape how agents think and what information they have access to.
+
+```python
+from crewai import Agent
+
+agent = Agent(
+    role="Security Auditor",
+    goal="Audit cloud infrastructure for vulnerabilities",
+    backstory="Expert in cloud security with 10 years of experience",
+    skills=["./skills/security-audit"],        # Domain instructions
+    knowledge_sources=[pdf_source, url_source], # Retrieved facts
+)
+```
+
+---
+
+## When to Use What
+
+| You need...                                       | Use               | Example                                  |
+| :------------------------------------------------ | :---------------- | :--------------------------------------- |
+| Agent to search the web                           | **Tools**         | `tools=[SerperDevTool()]`                |
+| Agent to call a remote API via MCP                | **MCPs**          | `mcps=["https://api.example.com/sse"]`   |
+| Agent to send emails via Gmail                    | **Apps**          | `apps=["gmail"]`                         |
+| Agent to follow specific procedures               | **Skills**        | `skills=["./skills/code-review"]`        |
+| Agent to reference company docs                   | **Knowledge**     | `knowledge_sources=[pdf_source]`         |
+| Agent to search the web AND follow review guidelines | **Tools + Skills** | Use both together                     |
+
+---
+
+## Combining Capabilities
+
+In practice, agents often use **multiple capability types together**. Here's a realistic example:
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+# A fully-equipped research agent
+researcher = Agent(
+    role="Senior Research Analyst",
+    goal="Produce comprehensive market analysis reports",
+    backstory="Expert analyst with deep industry knowledge",
+
+    # ACTION: What the agent can DO
+    tools=[
+        SerperDevTool(),         # Search the web
+        FileReadTool(),          # Read local files
+        CodeInterpreterTool(),   # Run Python code for analysis
+    ],
+    mcps=["https://data-api.example.com/sse"],  # Access remote data API
+    apps=["google_sheets"],                      # Write to Google Sheets
+
+    # CONTEXT: What the agent KNOWS
+    skills=["./skills/research-methodology"],    # How to conduct research
+    knowledge_sources=[company_docs],            # Company-specific data
+)
+```
+
+---
+
+## Comparison Table
+
+| Feature | Tools | MCPs | Apps | Skills | Knowledge |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **Gives agent actions** | ✅ | ✅ | ✅ | ❌ | ❌ |
+| **Modifies prompt** | ❌ | ❌ | ❌ | ✅ | ✅ |
+| **Requires code** | Yes | Config only | Config only | Markdown only | Config only |
+| **Runs locally** | Yes | Depends | Yes (with env var) | N/A | Yes |
+| **Needs API keys** | Per tool | Per server | Integration token | No | Embedder only |
+| **Set on Agent** | `tools=[]` | `mcps=[]` | `apps=[]` | `skills=[]` | `knowledge_sources=[]` |
+| **Set on Crew** | ❌ | ❌ | ❌ | `skills=[]` | `knowledge_sources=[]` |
+
+---
+
+## Deep Dives
+
+Ready to learn more about each capability type?
+
+<CardGroup cols={2}>
+  <Card title="Tools" icon="wrench" href="/en/concepts/tools">
+    Create custom tools, use the 75+ OSS catalog, configure caching and async execution.
+  </Card>
+  <Card title="MCP Integration" icon="plug" href="/en/mcp/overview">
+    Connect to MCP servers via stdio, SSE, or HTTP. Filter tools, configure auth.
+  </Card>
+  <Card title="Skills" icon="bolt" href="/en/concepts/skills">
+    Build skill packages with SKILL.md, inject domain expertise, use progressive disclosure.
+  </Card>
+  <Card title="Knowledge" icon="book" href="/en/concepts/knowledge">
+    Add knowledge from PDFs, CSVs, URLs, and more. Configure embedders and retrieval.
+  </Card>
+</CardGroup>

docs/en/concepts/skills.mdx

@@ -1,27 +1,186 @@
 ---
 title: Skills
-description: Filesystem-based skill packages that inject context into agent prompts.
+description: Filesystem-based skill packages that inject domain expertise and instructions into agent prompts.
 icon: bolt
 mode: "wide"
 ---
 
 ## Overview
 
-Skills are self-contained directories that provide agents with domain-specific instructions, references, and assets. Each skill is defined by a `SKILL.md` file with YAML frontmatter and a markdown body.
+Skills are self-contained directories that provide agents with **domain-specific instructions, guidelines, and reference material**. Each skill is defined by a `SKILL.md` file with YAML frontmatter and a markdown body.
 
-Skills use **progressive disclosure** — metadata is loaded first, full instructions only when activated, and resource catalogs only when needed.
+When activated, a skill's instructions are injected directly into the agent's task prompt — giving the agent expertise without requiring any code changes.
 
-## Directory Structure
+<Note type="info" title="Skills vs Tools — The Key Distinction">
+**Skills are NOT tools.** This is the most common point of confusion.
+
+- **Skills** inject *instructions and context* into the agent's prompt. They tell the agent *how to think* about a problem.
+- **Tools** give the agent *callable functions* to take action (search, read files, call APIs).
+
+You often need **both**: skills for expertise, tools for action. They are configured independently and complement each other.
+</Note>
+
+---
+
+## Quick Start
+
+### 1. Create a Skill Directory
 
 ```
-my-skill/
-├── SKILL.md            # Required — frontmatter + instructions
-├── scripts/            # Optional — executable scripts
-├── references/         # Optional — reference documents
-└── assets/             # Optional — static files (configs, data)
+skills/
+└── code-review/
+    ├── SKILL.md          # Required — instructions
+    ├── references/       # Optional — reference docs
+    │   └── style-guide.md
+    └── scripts/          # Optional — executable scripts
 ```
 
-The directory name must match the `name` field in `SKILL.md`.
+### 2. Write Your SKILL.md
+
+```markdown
+---
+name: code-review
+description: Guidelines for conducting thorough code reviews with focus on security and performance.
+metadata:
+  author: your-team
+  version: "1.0"
+---
+
+## Code Review Guidelines
+
+When reviewing code, follow this checklist:
+
+1. **Security**: Check for injection vulnerabilities, auth bypasses, and data exposure
+2. **Performance**: Look for N+1 queries, unnecessary allocations, and blocking calls
+3. **Readability**: Ensure clear naming, appropriate comments, and consistent style
+4. **Testing**: Verify adequate test coverage for new functionality
+
+### Severity Levels
+- **Critical**: Security vulnerabilities, data loss risks → block merge
+- **Major**: Performance issues, logic errors → request changes
+- **Minor**: Style issues, naming suggestions → approve with comments
+```
+
+### 3. Attach to an Agent
+
+```python
+from crewai import Agent
+from crewai_tools import GithubSearchTool, FileReadTool
+
+reviewer = Agent(
+    role="Senior Code Reviewer",
+    goal="Review pull requests for quality and security issues",
+    backstory="Staff engineer with expertise in secure coding practices.",
+    skills=["./skills"],                          # Injects review guidelines
+    tools=[GithubSearchTool(), FileReadTool()],   # Lets agent read code
+)
+```
+
+The agent now has both **expertise** (from the skill) and **capabilities** (from the tools).
+
+---
+
+## Skills + Tools: Working Together
+
+Here are common patterns showing how skills and tools complement each other:
+
+### Pattern 1: Skills Only (Domain Expertise, No Actions Needed)
+
+Use when the agent needs specific instructions but doesn't need to call external services:
+
+```python
+agent = Agent(
+    role="Technical Writer",
+    goal="Write clear API documentation",
+    backstory="Expert technical writer",
+    skills=["./skills/api-docs-style"],  # Writing guidelines and templates
+    # No tools needed — agent writes based on provided context
+)
+```
+
+### Pattern 2: Tools Only (Actions, No Special Expertise)
+
+Use when the agent needs to take action but doesn't need domain-specific instructions:
+
+```python
+from crewai_tools import SerperDevTool, ScrapeWebsiteTool
+
+agent = Agent(
+    role="Web Researcher",
+    goal="Find information about a topic",
+    backstory="Skilled at finding information online",
+    tools=[SerperDevTool(), ScrapeWebsiteTool()],  # Can search and scrape
+    # No skills needed — general research doesn't need special guidelines
+)
+```
+
+### Pattern 3: Skills + Tools (Expertise AND Actions)
+
+The most common real-world pattern. The skill provides *how* to approach the work; tools provide *what* the agent can do:
+
+```python
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+analyst = Agent(
+    role="Security Analyst",
+    goal="Audit infrastructure for vulnerabilities",
+    backstory="Expert in cloud security and compliance",
+    skills=["./skills/security-audit"],   # Audit methodology and checklists
+    tools=[
+        SerperDevTool(),                  # Research known vulnerabilities
+        FileReadTool(),                   # Read config files
+        CodeInterpreterTool(),            # Run analysis scripts
+    ],
+)
+```
+
+### Pattern 4: Skills + MCPs
+
+Skills work alongside MCP servers the same way they work with tools:
+
+```python
+agent = Agent(
+    role="Data Analyst",
+    goal="Analyze customer data and generate reports",
+    backstory="Expert data analyst with strong statistical background",
+    skills=["./skills/data-analysis"],           # Analysis methodology
+    mcps=["https://data-warehouse.example.com/sse"],  # Remote data access
+)
+```
+
+### Pattern 5: Skills + Apps
+
+Skills can guide how an agent uses platform integrations:
+
+```python
+agent = Agent(
+    role="Customer Support Agent",
+    goal="Respond to customer inquiries professionally",
+    backstory="Experienced support representative",
+    skills=["./skills/support-playbook"],  # Response templates and escalation rules
+    apps=["gmail", "zendesk"],             # Can send emails and update tickets
+)
+```
+
+---
+
+## Crew-Level Skills
+
+Skills can be set on a crew to apply to **all agents**:
+
+```python
+from crewai import Crew
+
+crew = Crew(
+    agents=[researcher, writer, reviewer],
+    tasks=[research_task, write_task, review_task],
+    skills=["./skills"],  # All agents get these skills
+)
+```
+
+Agent-level skills take priority — if the same skill is discovered at both levels, the agent's version is used.
+
+---
 
 ## SKILL.md Format
 
@@ -34,7 +193,7 @@ compatibility: crewai>=0.1.0          # optional
 metadata:                              # optional
   author: your-name
   version: "1.0"
-allowed-tools: web-search file-read   # optional, space-delimited
+allowed-tools: web-search file-read   # optional, experimental
 ---
 
 Instructions for the agent go here. This markdown body is injected
@@ -43,57 +202,46 @@ into the agent's prompt when the skill is activated.
 
 ### Frontmatter Fields
 
-| Field           | Required | Constraints                                                              |
+| Field           | Required | Description                                                              |
 | :-------------- | :------- | :----------------------------------------------------------------------- |
-| `name`          | Yes      | 1–64 chars. Lowercase alphanumeric and hyphens. No leading/trailing/consecutive hyphens. Must match directory name. |
+| `name`          | Yes      | 1–64 chars. Lowercase alphanumeric and hyphens. Must match directory name. |
 | `description`   | Yes      | 1–1024 chars. Describes what the skill does and when to use it.          |
 | `license`       | No       | License name or reference to a bundled license file.                     |
 | `compatibility` | No       | Max 500 chars. Environment requirements (products, packages, network).   |
 | `metadata`      | No       | Arbitrary string key-value mapping.                                      |
 | `allowed-tools` | No       | Space-delimited list of pre-approved tools. Experimental.                |
 
-## Usage
+---
 
-### Agent-level Skills
+## Directory Structure
 
-Pass skill directory paths to an agent:
-
-```python
-from crewai import Agent
-
-agent = Agent(
-    role="Researcher",
-    goal="Find relevant information",
-    backstory="An expert researcher.",
-    skills=["./skills"],  # discovers all skills in this directory
-)
+```
+my-skill/
+├── SKILL.md            # Required — frontmatter + instructions
+├── scripts/            # Optional — executable scripts
+├── references/         # Optional — reference documents
+└── assets/             # Optional — static files (configs, data)
 ```
 
-### Crew-level Skills
+The directory name must match the `name` field in `SKILL.md`. The `scripts/`, `references/`, and `assets/` directories are available on the skill's `path` for agents that need to reference files directly.
 
-Skill paths on a crew are merged into every agent:
+---
 
-```python
-from crewai import Crew
+## Pre-loading Skills
 
-crew = Crew(
-    agents=[agent],
-    tasks=[task],
-    skills=["./skills"],
-)
-```
-
-### Pre-loaded Skills
-
-You can also pass `Skill` objects directly:
+For more control, you can discover and activate skills programmatically:
 
 ```python
 from pathlib import Path
 from crewai.skills import discover_skills, activate_skill
 
+# Discover all skills in a directory
 skills = discover_skills(Path("./skills"))
+
+# Activate them (loads full SKILL.md body)
 activated = [activate_skill(s) for s in skills]
 
+# Pass to an agent
 agent = Agent(
     role="Researcher",
     goal="Find relevant information",
@@ -102,14 +250,57 @@ agent = Agent(
 )
 ```
 
+---
+
 ## How Skills Are Loaded
 
-Skills load progressively — only the data needed at each stage is read:
+Skills use **progressive disclosure** — only loading what's needed at each stage:
 
-| Stage            | What's loaded                                     | When               |
-| :--------------- | :------------------------------------------------ | :----------------- |
-| Discovery        | Name, description, frontmatter fields             | `discover_skills()` |
-| Activation       | Full SKILL.md body text                           | `activate_skill()` |
+| Stage      | What's loaded                         | When                |
+| :--------- | :------------------------------------ | :------------------ |
+| Discovery  | Name, description, frontmatter fields | `discover_skills()` |
+| Activation | Full SKILL.md body text               | `activate_skill()`  |
 
-During normal agent execution, skills are automatically discovered and activated. The `scripts/`, `references/`, and `assets/` directories are available on the skill's `path` for agents that need to reference files directly.
+During normal agent execution (passing directory paths via `skills=["./skills"]`), skills are automatically discovered and activated. The progressive loading only matters when using the programmatic API.
 
+---
+
+## Skills vs Knowledge
+
+Both skills and knowledge modify the agent's prompt, but they serve different purposes:
+
+| Aspect | Skills | Knowledge |
+| :--- | :--- | :--- |
+| **What it provides** | Instructions, procedures, guidelines | Facts, data, information |
+| **How it's stored** | Markdown files (SKILL.md) | Embedded in vector store (ChromaDB) |
+| **How it's retrieved** | Entire body injected into prompt | Semantic search finds relevant chunks |
+| **Best for** | Methodology, checklists, style guides | Company docs, product info, reference data |
+| **Set via** | `skills=["./skills"]` | `knowledge_sources=[source]` |
+
+**Rule of thumb:** If the agent needs to follow a *process*, use a skill. If the agent needs to reference *data*, use knowledge.
+
+---
+
+## Common Questions
+
+<AccordionGroup>
+  <Accordion title="Do I need to set skills AND tools?">
+    It depends on your use case. Skills and tools are **independent** — you can use either, both, or neither.
+
+    - **Skills alone**: When the agent needs expertise but no external actions (e.g., writing with style guidelines)
+    - **Tools alone**: When the agent needs actions but no special methodology (e.g., simple web search)
+    - **Both**: When the agent needs expertise AND actions (e.g., security audit with specific checklists AND ability to scan code)
+  </Accordion>
+
+  <Accordion title="Do skills automatically provide tools?">
+    **No.** The `allowed-tools` field in SKILL.md is experimental metadata only — it does not provision or inject any tools. You must always set tools separately via `tools=[]`, `mcps=[]`, or `apps=[]`.
+  </Accordion>
+
+  <Accordion title="What happens if I set the same skill on both an agent and its crew?">
+    The agent-level skill takes priority. Skills are deduplicated by name — the agent's skills are processed first, so if the same skill name appears at both levels, the agent's version is used.
+  </Accordion>
+
+  <Accordion title="How large can a SKILL.md body be?">
+    There's a soft warning at 50,000 characters, but no hard limit. Keep skills focused and concise for best results — large prompt injections can dilute the agent's attention.
+  </Accordion>
+</AccordionGroup>

docs/en/concepts/tools.mdx

@@ -10,6 +10,10 @@ mode: "wide"
 CrewAI tools empower agents with capabilities ranging from web searching and data analysis to collaboration and delegating tasks among coworkers.
 This documentation outlines how to create, integrate, and leverage these tools within the CrewAI framework, including a new focus on collaboration tools.
 
+<Note type="info" title="Tools are one of five agent capability types">
+  Tools give agents **callable functions** to take action. They work alongside [MCPs](/en/mcp/overview) (remote tool servers), [Apps](/en/concepts/agent-capabilities) (platform integrations), [Skills](/en/concepts/skills) (domain expertise), and [Knowledge](/en/concepts/knowledge) (retrieved facts). See the [Agent Capabilities](/en/concepts/agent-capabilities) overview to understand when to use each.
+</Note>
+
 ## What is a Tool?
 
 A tool in CrewAI is a skill or function that agents can utilize to perform various actions.

docs/en/enterprise/features/sso.mdx

@@ -0,0 +1,550 @@
+---
+title: Single Sign-On (SSO)
+icon: "key"
+description: Configure enterprise SSO authentication for CrewAI Platform — SaaS and Factory
+---
+
+## Overview
+
+CrewAI Platform supports enterprise Single Sign-On (SSO) across both **SaaS (AMP)** and **Factory (self-hosted)** deployments. SSO enables your team to authenticate using your organization's existing identity provider, enforcing centralized access control, MFA policies, and user lifecycle management.
+
+### Supported Providers
+
+| Provider | SaaS | Factory | Protocol | CLI Support |
+|---|---|---|---|---|
+| **WorkOS** | ✅ (default) | ✅ | OAuth 2.0 / OIDC | ✅ |
+| **Microsoft Entra ID** (Azure AD) | ✅ (enterprise) | ✅ | OAuth 2.0 / SAML 2.0 | ✅ |
+| **Okta** | ✅ (enterprise) | ✅ | OAuth 2.0 / OIDC | ✅ |
+| **Auth0** | ✅ (enterprise) | ✅ | OAuth 2.0 / OIDC | ✅ |
+| **Keycloak** | — | ✅ | OAuth 2.0 / OIDC | ✅ |
+
+### Key Capabilities
+
+- **SAML 2.0 and OAuth 2.0 / OIDC** protocol support
+- **Device Authorization Grant** flow for CLI authentication
+- **Role-Based Access Control (RBAC)** with custom roles and per-resource permissions
+- **MFA enforcement** delegated to your identity provider
+- **User provisioning** through IdP assignment (users/groups)
+
+---
+
+## SaaS SSO
+
+### Default Authentication
+
+CrewAI's managed SaaS platform (AMP) uses **WorkOS** as the default authentication provider. When you sign up at [app.crewai.com](https://app.crewai.com), authentication is handled through `login.crewai.com` — no additional SSO configuration is required.
+
+### Enterprise Custom SSO
+
+Enterprise SaaS customers can configure SSO with their own identity provider (Entra ID, Okta, Auth0). Contact your CrewAI account team to enable custom SSO for your organization. Once configured:
+
+1. Your team members authenticate through your organization's IdP
+2. Access control and MFA policies are enforced by your IdP
+3. The CrewAI CLI automatically detects your SSO configuration via `crewai enterprise configure`
+
+### CLI Defaults (SaaS)
+
+| Setting | Default Value |
+|---|---|
+| `enterprise_base_url` | `https://app.crewai.com` |
+| `oauth2_provider` | `workos` |
+| `oauth2_domain` | `login.crewai.com` |
+
+---
+
+## Factory SSO Setup
+
+Factory (self-hosted) deployments require you to configure SSO by setting environment variables in your Helm `values.yaml` and registering an application in your identity provider.
+
+### Microsoft Entra ID (Azure AD)
+
+<Steps>
+  <Step title="Register an Application">
+    1. Go to [portal.azure.com](https://portal.azure.com) → **Microsoft Entra ID** → **App registrations** → **New registration**
+    2. Configure:
+       - **Name:** `CrewAI` (or your preferred name)
+       - **Supported account types:** Accounts in this organizational directory only
+       - **Redirect URI:** Select **Web**, enter `https://<your-domain>/auth/entra_id/callback`
+    3. Click **Register**
+  </Step>
+
+  <Step title="Collect Credentials">
+    From the app overview page, copy:
+    - **Application (client) ID** → `ENTRA_ID_CLIENT_ID`
+    - **Directory (tenant) ID** → `ENTRA_ID_TENANT_ID`
+  </Step>
+
+  <Step title="Create Client Secret">
+    1. Navigate to **Certificates & Secrets** → **New client secret**
+    2. Add a description and select expiration period
+    3. Copy the secret value immediately (it won't be shown again) → `ENTRA_ID_CLIENT_SECRET`
+  </Step>
+
+  <Step title="Grant Admin Consent">
+    1. Go to **Enterprise applications** → select your app
+    2. Under **Security** → **Permissions**, click **Grant admin consent**
+    3. Ensure **Microsoft Graph → User.Read** is granted
+  </Step>
+
+  <Step title="Configure App Roles (Recommended)">
+    Under **App registrations** → your app → **App roles**, create:
+
+    | Display Name | Value | Allowed Member Types |
+    |---|---|---|
+    | Member | `member` | Users/Groups |
+    | Factory Admin | `factory-admin` | Users/Groups |
+
+    <Note>
+      The `member` role grants login access. The `factory-admin` role grants admin panel access. Roles are included in the JWT automatically.
+    </Note>
+  </Step>
+
+  <Step title="Assign Users">
+    1. Under **Properties**, set **Assignment required?** to **Yes**
+    2. Under **Users and groups**, assign users/groups with the appropriate role
+  </Step>
+
+  <Step title="Set Environment Variables">
+    ```yaml
+    envVars:
+      AUTH_PROVIDER: "entra_id"
+
+    secrets:
+      ENTRA_ID_CLIENT_ID: "<Application (client) ID>"
+      ENTRA_ID_CLIENT_SECRET: "<Client Secret>"
+      ENTRA_ID_TENANT_ID: "<Directory (tenant) ID>"
+    ```
+  </Step>
+
+  <Step title="Enable CLI Support (Optional)">
+    To allow `crewai login` via Device Authorization Grant:
+
+    1. Under **Authentication** → **Advanced settings**, enable **Allow public client flows**
+    2. Under **Expose an API**, add an Application ID URI (e.g., `api://crewai-cli`)
+    3. Add a scope (e.g., `read`) with **Admins and users** consent
+    4. Under **Manifest**, set `accessTokenAcceptedVersion` to `2`
+    5. Add environment variables:
+
+    ```yaml
+    secrets:
+      ENTRA_ID_DEVICE_AUTHORIZATION_CLIENT_ID: "<Application (client) ID>"
+      ENTRA_ID_CUSTOM_OPENID_SCOPE: "<scope URI, e.g. api://crewai-cli/read>"
+    ```
+  </Step>
+</Steps>
+
+---
+
+### Okta
+
+<Steps>
+  <Step title="Create App Integration">
+    1. Open Okta Admin Console → **Applications** → **Create App Integration**
+    2. Select **OIDC - OpenID Connect** → **Web Application** → **Next**
+    3. Configure:
+       - **App integration name:** `CrewAI SSO`
+       - **Sign-in redirect URI:** `https://<your-domain>/auth/okta/callback`
+       - **Sign-out redirect URI:** `https://<your-domain>`
+       - **Assignments:** Choose who can access (everyone or specific groups)
+    4. Click **Save**
+  </Step>
+
+  <Step title="Collect Credentials">
+    From the app details page:
+    - **Client ID** → `OKTA_CLIENT_ID`
+    - **Client Secret** → `OKTA_CLIENT_SECRET`
+    - **Okta URL** (top-right corner, under your username) → `OKTA_SITE`
+  </Step>
+
+  <Step title="Configure Authorization Server">
+    1. Navigate to **Security** → **API**
+    2. Select your authorization server (default: `default`)
+    3. Under **Access Policies**, add a policy and rule:
+       - In the rule, under **Scopes requested**, select **The following scopes** → **OIDC default scopes**
+    4. Note the **Name** and **Audience** of the authorization server
+
+    <Warning>
+      The authorization server name and audience must match `OKTA_AUTHORIZATION_SERVER` and `OKTA_AUDIENCE` exactly. Mismatches cause `401 Unauthorized` or `Invalid token: Signature verification failed` errors.
+    </Warning>
+  </Step>
+
+  <Step title="Set Environment Variables">
+    ```yaml
+    envVars:
+      AUTH_PROVIDER: "okta"
+
+    secrets:
+      OKTA_CLIENT_ID: "<Okta app client ID>"
+      OKTA_CLIENT_SECRET: "<Okta client secret>"
+      OKTA_SITE: "https://your-domain.okta.com"
+      OKTA_AUTHORIZATION_SERVER: "default"
+      OKTA_AUDIENCE: "api://default"
+    ```
+  </Step>
+
+  <Step title="Enable CLI Support (Optional)">
+    1. Create a **new** app integration: **OIDC** → **Native Application**
+    2. Enable **Device Authorization** and **Refresh Token** grant types
+    3. Allow everyone in your organization to access
+    4. Add environment variable:
+
+    ```yaml
+    secrets:
+      OKTA_DEVICE_AUTHORIZATION_CLIENT_ID: "<Native app client ID>"
+    ```
+
+    <Note>
+      Device Authorization requires a **Native Application** — it cannot use the Web Application created for browser-based SSO.
+    </Note>
+  </Step>
+</Steps>
+
+---
+
+### Keycloak
+
+<Steps>
+  <Step title="Create a Client">
+    1. Open Keycloak Admin Console → navigate to your realm
+    2. **Clients** → **Create client**:
+       - **Client type:** OpenID Connect
+       - **Client ID:** `crewai-factory` (suggested)
+    3. Capability config:
+       - **Client authentication:** On
+       - **Standard flow:** Checked
+    4. Login settings:
+       - **Root URL:** `https://<your-domain>`
+       - **Valid redirect URIs:** `https://<your-domain>/auth/keycloak/callback`
+       - **Valid post logout redirect URIs:** `https://<your-domain>`
+    5. Click **Save**
+  </Step>
+
+  <Step title="Collect Credentials">
+    - **Client ID** → `KEYCLOAK_CLIENT_ID`
+    - Under **Credentials** tab: **Client secret** → `KEYCLOAK_CLIENT_SECRET`
+    - **Realm name** → `KEYCLOAK_REALM`
+    - **Keycloak server URL** → `KEYCLOAK_SITE`
+  </Step>
+
+  <Step title="Set Environment Variables">
+    ```yaml
+    envVars:
+      AUTH_PROVIDER: "keycloak"
+
+    secrets:
+      KEYCLOAK_CLIENT_ID: "<client ID>"
+      KEYCLOAK_CLIENT_SECRET: "<client secret>"
+      KEYCLOAK_SITE: "https://keycloak.yourdomain.com"
+      KEYCLOAK_REALM: "<realm name>"
+      KEYCLOAK_AUDIENCE: "account"
+      # Only set if using a custom base path (pre-v17 migrations):
+      # KEYCLOAK_BASE_URL: "/auth"
+    ```
+
+    <Note>
+      Keycloak includes `account` as the default audience in access tokens. For most installations, `KEYCLOAK_AUDIENCE=account` works without additional configuration. See [Keycloak audience documentation](https://www.keycloak.org/docs/latest/authorization_services/index.html) if you need a custom audience.
+    </Note>
+  </Step>
+
+  <Step title="Enable CLI Support (Optional)">
+    1. Create a **second** client:
+       - **Client type:** OpenID Connect
+       - **Client ID:** `crewai-factory-cli` (suggested)
+       - **Client authentication:** Off (Device Authorization requires a public client)
+       - **Authentication flow:** Check **only** OAuth 2.0 Device Authorization Grant
+    2. Add environment variable:
+
+    ```yaml
+    secrets:
+      KEYCLOAK_DEVICE_AUTHORIZATION_CLIENT_ID: "<CLI client ID>"
+    ```
+  </Step>
+</Steps>
+
+---
+
+### WorkOS
+
+<Steps>
+  <Step title="Configure in WorkOS Dashboard">
+    1. Create an application in the [WorkOS Dashboard](https://dashboard.workos.com)
+    2. Configure the redirect URI: `https://<your-domain>/auth/workos/callback`
+    3. Note the **Client ID** and **AuthKit domain**
+    4. Set up organizations in the WorkOS dashboard
+  </Step>
+
+  <Step title="Set Environment Variables">
+    ```yaml
+    envVars:
+      AUTH_PROVIDER: "workos"
+
+    secrets:
+      WORKOS_CLIENT_ID: "<WorkOS client ID>"
+      WORKOS_AUTHKIT_DOMAIN: "<your-authkit-domain.authkit.com>"
+    ```
+  </Step>
+</Steps>
+
+---
+
+### Auth0
+
+<Steps>
+  <Step title="Create Application">
+    1. In the [Auth0 Dashboard](https://manage.auth0.com), create a new **Regular Web Application**
+    2. Configure:
+       - **Allowed Callback URLs:** `https://<your-domain>/auth/auth0/callback`
+       - **Allowed Logout URLs:** `https://<your-domain>`
+    3. Note the **Domain**, **Client ID**, and **Client Secret**
+  </Step>
+
+  <Step title="Set Environment Variables">
+    ```yaml
+    envVars:
+      AUTH_PROVIDER: "auth0"
+
+    secrets:
+      AUTH0_CLIENT_ID: "<Auth0 client ID>"
+      AUTH0_CLIENT_SECRET: "<Auth0 client secret>"
+      AUTH0_DOMAIN: "<your-tenant.auth0.com>"
+    ```
+  </Step>
+
+  <Step title="Enable CLI Support (Optional)">
+    1. Create a **Native** application in Auth0 for Device Authorization
+    2. Enable the **Device Authorization** grant type under application settings
+    3. Configure the CLI with the appropriate audience and client ID
+  </Step>
+</Steps>
+
+---
+
+## CLI Authentication
+
+The CrewAI CLI supports SSO authentication via the **Device Authorization Grant** flow. This allows developers to authenticate from their terminal without exposing credentials.
+
+### Quick Setup
+
+For Factory installations, the CLI can auto-configure all OAuth2 settings:
+
+```bash
+crewai enterprise configure https://your-factory-url.app
+```
+
+This command fetches the SSO configuration from your Factory instance and sets all required CLI parameters automatically.
+
+Then authenticate:
+
+```bash
+crewai login
+```
+
+<Note>
+  Requires CrewAI CLI version **1.6.0** or higher for Entra ID, **0.159.0** or higher for Okta, and **1.9.0** or higher for Keycloak.
+</Note>
+
+### Manual CLI Configuration
+
+If you need to configure the CLI manually, use `crewai config set`:
+
+```bash
+# Set the provider
+crewai config set oauth2_provider okta
+
+# Set provider-specific values
+crewai config set oauth2_domain your-domain.okta.com
+crewai config set oauth2_client_id your-client-id
+crewai config set oauth2_audience api://default
+
+# Set the enterprise base URL
+crewai config set enterprise_base_url https://your-factory-url.app
+```
+
+### CLI Configuration Reference
+
+| Setting | Description | Example |
+|---|---|---|
+| `enterprise_base_url` | Your CrewAI instance URL | `https://crewai.yourcompany.com` |
+| `oauth2_provider` | Provider name | `workos`, `okta`, `auth0`, `entra_id`, `keycloak` |
+| `oauth2_domain` | Provider domain | `your-domain.okta.com` |
+| `oauth2_client_id` | OAuth2 client ID | `0oaqnwji7pGW7VT6T697` |
+| `oauth2_audience` | API audience identifier | `api://default` |
+
+View current configuration:
+
+```bash
+crewai config list
+```
+
+### How Device Authorization Works
+
+1. Run `crewai login` — the CLI requests a device code from your IdP
+2. A verification URL and code are displayed in your terminal
+3. Your browser opens to the verification URL
+4. Enter the code and authenticate with your IdP credentials
+5. The CLI receives an access token and stores it locally
+
+---
+
+## Role-Based Access Control (RBAC)
+
+CrewAI Platform provides granular RBAC that integrates with your SSO provider.
+
+### Permission Model
+
+| Permission | Description |
+|---|---|
+| **Read** | View resources (dashboards, automations, logs) |
+| **Write** | Create and modify resources |
+| **Manage** | Full control including deletion and configuration |
+
+### Resources
+
+Permissions can be scoped to individual resources:
+
+- **Usage Dashboard** — Platform usage metrics and analytics
+- **Automations Dashboard** — Crew and flow management
+- **Environment Variables** — Secret and configuration management
+- **Individual Automations** — Per-automation access control
+
+### Roles
+
+- **Predefined roles** come out of the box with standard permission sets
+- **Custom roles** can be created with any combination of permissions
+- **Per-resource assignment** — limit specific automations to individual users or roles
+
+### Factory Admin Access
+
+For Factory deployments using Entra ID, admin access is controlled via App Roles:
+
+- Assign the `factory-admin` role to users who need admin panel access
+- Assign the `member` role for standard platform access
+- Roles are communicated via JWT claims — no additional configuration needed after IdP setup
+
+---
+
+## Troubleshooting
+
+### Invalid Redirect URI
+
+**Symptom:** Authentication fails with a redirect URI mismatch error.
+
+**Fix:** Ensure the redirect URI in your IdP exactly matches the expected callback URL:
+
+| Provider | Callback URL |
+|---|---|
+| Entra ID | `https://<domain>/auth/entra_id/callback` |
+| Okta | `https://<domain>/auth/okta/callback` |
+| Keycloak | `https://<domain>/auth/keycloak/callback` |
+| WorkOS | `https://<domain>/auth/workos/callback` |
+| Auth0 | `https://<domain>/auth/auth0/callback` |
+
+### CLI Login Fails (Device Authorization)
+
+**Symptom:** `crewai login` returns an error or times out.
+
+**Fix:**
+- Verify that Device Authorization Grant is enabled in your IdP
+- For Okta: ensure you have a **Native Application** (not Web) with Device Authorization grant
+- For Entra ID: ensure **Allow public client flows** is enabled
+- For Keycloak: ensure the CLI client has **Client authentication: Off** and only Device Authorization Grant enabled
+- Check that `*_DEVICE_AUTHORIZATION_CLIENT_ID` environment variable is set on the server
+
+### Token Validation Errors
+
+**Symptom:** `Invalid token: Signature verification failed` or `401 Unauthorized` after login.
+
+**Fix:**
+- **Okta:** Verify `OKTA_AUTHORIZATION_SERVER` and `OKTA_AUDIENCE` match the authorization server's Name and Audience exactly
+- **Entra ID:** Ensure `accessTokenAcceptedVersion` is set to `2` in the app manifest
+- **Keycloak:** Verify `KEYCLOAK_AUDIENCE` matches the audience in your access tokens (default: `account`)
+
+### Admin Consent Not Granted (Entra ID)
+
+**Symptom:** Users can't log in, see "needs admin approval" message.
+
+**Fix:** Go to **Enterprise applications** → your app → **Permissions** → **Grant admin consent**. Ensure `User.Read` is granted for Microsoft Graph.
+
+### 403 Forbidden After Login
+
+**Symptom:** User authenticates successfully but gets 403 errors.
+
+**Fix:**
+- Check that the user is assigned to the application in your IdP
+- For Entra ID with **Assignment required = Yes**: ensure the user has a role assignment (Member or Factory Admin)
+- For Okta: verify the user or their group is assigned under the app's **Assignments** tab
+
+### CLI Can't Reach Factory Instance
+
+**Symptom:** `crewai enterprise configure` fails to connect.
+
+**Fix:**
+- Verify the Factory URL is reachable from your machine
+- Check that `enterprise_base_url` is set correctly: `crewai config list`
+- Ensure TLS certificates are valid and trusted
+
+---
+
+## Environment Variables Reference
+
+### Common
+
+| Variable | Description |
+|---|---|
+| `AUTH_PROVIDER` | Authentication provider: `entra_id`, `okta`, `workos`, `auth0`, `keycloak`, `local` |
+
+### Microsoft Entra ID
+
+| Variable | Required | Description |
+|---|---|---|
+| `ENTRA_ID_CLIENT_ID` | ✅ | Application (client) ID from Azure |
+| `ENTRA_ID_CLIENT_SECRET` | ✅ | Client secret from Azure |
+| `ENTRA_ID_TENANT_ID` | ✅ | Directory (tenant) ID from Azure |
+| `ENTRA_ID_DEVICE_AUTHORIZATION_CLIENT_ID` | CLI only | Client ID for Device Authorization Grant |
+| `ENTRA_ID_CUSTOM_OPENID_SCOPE` | CLI only | Custom scope from "Expose an API" (e.g., `api://crewai-cli/read`) |
+
+### Okta
+
+| Variable | Required | Description |
+|---|---|---|
+| `OKTA_CLIENT_ID` | ✅ | Okta application client ID |
+| `OKTA_CLIENT_SECRET` | ✅ | Okta client secret |
+| `OKTA_SITE` | ✅ | Okta organization URL (e.g., `https://your-domain.okta.com`) |
+| `OKTA_AUTHORIZATION_SERVER` | ✅ | Authorization server name (e.g., `default`) |
+| `OKTA_AUDIENCE` | ✅ | Authorization server audience (e.g., `api://default`) |
+| `OKTA_DEVICE_AUTHORIZATION_CLIENT_ID` | CLI only | Native app client ID for Device Authorization |
+
+### WorkOS
+
+| Variable | Required | Description |
+|---|---|---|
+| `WORKOS_CLIENT_ID` | ✅ | WorkOS application client ID |
+| `WORKOS_AUTHKIT_DOMAIN` | ✅ | AuthKit domain (e.g., `your-domain.authkit.com`) |
+
+### Auth0
+
+| Variable | Required | Description |
+|---|---|---|
+| `AUTH0_CLIENT_ID` | ✅ | Auth0 application client ID |
+| `AUTH0_CLIENT_SECRET` | ✅ | Auth0 client secret |
+| `AUTH0_DOMAIN` | ✅ | Auth0 tenant domain (e.g., `your-tenant.auth0.com`) |
+
+### Keycloak
+
+| Variable | Required | Description |
+|---|---|---|
+| `KEYCLOAK_CLIENT_ID` | ✅ | Keycloak client ID |
+| `KEYCLOAK_CLIENT_SECRET` | ✅ | Keycloak client secret |
+| `KEYCLOAK_SITE` | ✅ | Keycloak server URL |
+| `KEYCLOAK_REALM` | ✅ | Keycloak realm name |
+| `KEYCLOAK_AUDIENCE` | ✅ | Token audience (default: `account`) |
+| `KEYCLOAK_BASE_URL` | Optional | Base URL path (e.g., `/auth` for pre-v17 migrations) |
+| `KEYCLOAK_DEVICE_AUTHORIZATION_CLIENT_ID` | CLI only | Public client ID for Device Authorization |
+
+---
+
+## Next Steps
+
+- [Installation Guide](/installation) — Get started with CrewAI
+- [Quickstart](/quickstart) — Build your first crew
+- [RBAC Setup](/enterprise/features/rbac) — Detailed role and permission management

docs/en/mcp/security.mdx

@@ -156,19 +156,7 @@ If you are developing an MCP server that CrewAI agents might connect to, conside
 - **Adherence to MCP Authorization Spec**: If implementing authentication and authorization, strictly follow the [MCP Authorization specification](https://modelcontextprotocol.io/specification/draft/basic/authorization) and relevant [OAuth 2.0 security best practices](https://datatracker.ietf.org/doc/html/rfc9700).
 - **Regular Security Audits**: If your MCP server handles sensitive data, performs critical operations, or is publicly exposed, consider periodic security audits by qualified professionals.
 
-## 5. Reporting Security Vulnerabilities
-
-If you discover a security vulnerability in CrewAI, please report it responsibly through our Bugcrowd Vulnerability Disclosure Program (VDP):
-
-**Submit reports to:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-**Do not** disclose vulnerabilities via public GitHub issues, pull requests, or social media. Reports submitted via channels other than Bugcrowd will not be reviewed.
-</Warning>
-
-For full details, see our [Security Policy](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md).
-
-## 6. Further Reading
+## 5. Further Reading
 
 For more detailed information on MCP security, refer to the official documentation:
 - **[MCP Transport Security](https://modelcontextprotocol.io/docs/concepts/transports#security-considerations)**

docs/en/security.mdx

@@ -1,22 +0,0 @@
----
-title: Security Policy
-description: Learn how to report security vulnerabilities and about CrewAI's security practices.
-icon: shield
-mode: "wide"
----
-
-## Reporting Security Vulnerabilities
-
-If you discover a security vulnerability in CrewAI, please report it responsibly through our Bugcrowd Vulnerability Disclosure Program (VDP):
-
-**Submit reports to:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-**Do not** disclose vulnerabilities via public GitHub issues, pull requests, or social media. Reports submitted via channels other than Bugcrowd will not be reviewed.
-</Warning>
-
-For full details, see our [Security Policy on GitHub](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md).
-
-## Security Resources
-
-- **[MCP Security Considerations](/mcp/security)** — Best practices for securely integrating MCP servers with your CrewAI agents, including transport security, prompt injection risks, and server implementation advice.

docs/ko/changelog.mdx

@@ -4,6 +4,127 @@ description: "CrewAI의 제품 업데이트, 개선 사항 및 버그 수정"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="2026년 3월 31일">
+  ## v1.13.0a5
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a5)
+
+  ## 변경 사항
+
+  ### 문서
+  - v1.13.0a4에 대한 변경 로그 및 버전 업데이트
+
+  ## 기여자
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="2026년 4월 1일">
+  ## v1.13.0a4
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a4)
+
+  ## 변경 사항
+
+  ### 문서
+  - v1.13.0a3에 대한 변경 로그 및 버전 업데이트
+
+  ## 기여자
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="2026년 4월 1일">
+  ## v1.13.0a3
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a3)
+
+  ## 변경 사항
+
+  ### 기능
+  - LLMCallCompletedEvent에서 토큰 사용 데이터 발행
+  - 도구 메타데이터를 AMP로 추출 및 게시
+
+  ### 버그 수정
+  - `stop` API 매개변수를 지원하지 않는 GPT-5.x 모델 처리
+
+  ### 문서
+  - 모든 언어에서 에이전트 기능의 부정확성 수정
+  - 에이전트 기능 개요 추가 및 기술 문서 개선
+  - 포괄적인 SSO 구성 가이드 추가
+  - v1.13.0rc1에 대한 변경 로그 및 버전 업데이트
+
+  ### 리팩토링
+  - Flow를 Pydantic BaseModel로 변환
+  - LLM 클래스를 Pydantic BaseModel로 변환
+  - InstanceOf[T]를 일반 타입 주석으로 교체
+  - 사용되지 않는 메서드 제거
+
+  ## 기여자
+
+  @dependabot[bot], @greysonlalonde, @iris-clawd, @lorenzejay, @lucasgomide, @thiagomoretto
+
+</Update>
+
+<Update label="2026년 3월 27일">
+  ## v1.13.0rc1
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0rc1)
+
+  ## 변경 사항
+
+  ### 문서
+  - v1.13.0a2의 변경 로그 및 버전 업데이트
+
+  ## 기여자
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="2026년 3월 27일">
+  ## v1.13.0a2
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a2)
+
+  ## 변경 사항
+
+  ### 기능
+  - 릴리스 중 자동 업데이트 배포 테스트 리포지토리
+  - 기업 릴리스의 복원력 및 사용자 경험 개선
+
+  ### 문서
+  - v1.13.0a1에 대한 변경 로그 및 버전 업데이트
+
+  ## 기여자
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="2026년 3월 27일">
+  ## v1.13.0a1
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a1)
+
+  ## 변경 사항
+
+  ### 버그 수정
+  - Node를 LTS 22로 고정하여 문서 작업 흐름의 끊어진 링크 수정
+  - 기업 릴리스에서 새로 게시된 패키지의 uv 캐시 초기화
+
+  ### 문서
+  - 포괄적인 RBAC 권한 매트릭스 및 배포 가이드 추가
+  - v1.12.2에 대한 변경 로그 및 버전 업데이트
+
+  ## 기여자
+
+  @greysonlalonde, @iris-clawd, @joaomdmoura
+
+</Update>
+
 <Update label="2026년 3월 25일">
   ## v1.12.2
 

docs/ko/concepts/agent-capabilities.mdx

@@ -0,0 +1,147 @@
+---
+title: "에이전트 기능"
+description: "CrewAI 에이전트를 확장하는 다섯 가지 방법 이해하기: 도구, MCP, 앱, 스킬, 지식."
+icon: puzzle-piece
+mode: "wide"
+---
+
+## 개요
+
+CrewAI 에이전트는 **다섯 가지 고유한 기능 유형**으로 확장할 수 있으며, 각각 다른 목적을 가지고 있습니다. 각 유형을 언제 사용해야 하는지, 그리고 어떻게 함께 작동하는지 이해하는 것이 효과적인 에이전트를 구축하는 핵심입니다.
+
+<CardGroup cols={2}>
+  <Card title="도구" icon="wrench" href="/ko/concepts/tools" color="#3B82F6">
+    **호출 가능한 함수** — 에이전트가 행동을 취할 수 있게 합니다. 웹 검색, 파일 작업, API 호출, 코드 실행.
+  </Card>
+  <Card title="MCP 서버" icon="plug" href="/ko/mcp/overview" color="#8B5CF6">
+    **원격 도구 서버** — Model Context Protocol을 통해 에이전트를 외부 도구 서버에 연결합니다. 도구와 같은 효과이지만 외부에서 호스팅됩니다.
+  </Card>
+  <Card title="앱" icon="grid-2" color="#EC4899">
+    **플랫폼 통합** — CrewAI 플랫폼을 통해 에이전트를 SaaS 앱(Gmail, Slack, Jira, Salesforce)에 연결합니다. 플랫폼 통합 토큰으로 로컬에서 실행됩니다.
+  </Card>
+  <Card title="스킬" icon="bolt" href="/ko/concepts/skills" color="#F59E0B">
+    **도메인 전문성** — 에이전트 프롬프트에 지침, 가이드라인 및 참조 자료를 주입합니다. 스킬은 에이전트에게 *어떻게 생각할지*를 알려줍니다.
+  </Card>
+  <Card title="지식" icon="book" href="/ko/concepts/knowledge" color="#10B981">
+    **검색된 사실** — 시맨틱 검색(RAG)을 통해 문서, 파일 및 URL에서 에이전트에게 데이터를 제공합니다. 지식은 에이전트에게 *무엇을 알아야 하는지*를 제공합니다.
+  </Card>
+</CardGroup>
+
+---
+
+## 핵심 구분
+
+가장 중요한 점: **이 기능들은 두 가지 범주로 나뉩니다**.
+
+### 액션 기능 (도구, MCP, 앱)
+
+에이전트에게 **무언가를 할 수 있는** 능력을 부여합니다 — API 호출, 파일 읽기, 웹 검색, 이메일 전송. 실행 시점에 세 가지 모두 동일한 내부 형식(`BaseTool` 인스턴스)으로 변환되며, 에이전트가 호출할 수 있는 통합 도구 목록에 나타납니다.
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool
+
+agent = Agent(
+    role="Researcher",
+    goal="Find and compile market data",
+    backstory="Expert market analyst",
+    tools=[SerperDevTool(), FileReadTool()],  # 로컬 도구
+    mcps=["https://mcp.example.com/sse"],     # 원격 MCP 서버 도구
+    apps=["gmail", "google_sheets"],           # 플랫폼 통합
+)
+```
+
+### 컨텍스트 기능 (스킬, 지식)
+
+에이전트의 **프롬프트**를 수정합니다 — 에이전트가 추론을 시작하기 전에 전문성, 지침 또는 검색된 데이터를 주입합니다. 에이전트에게 새로운 액션을 제공하는 것이 아니라, 에이전트가 어떻게 생각하고 어떤 정보에 접근할 수 있는지를 형성합니다.
+
+```python
+from crewai import Agent
+
+agent = Agent(
+    role="Security Auditor",
+    goal="Audit cloud infrastructure for vulnerabilities",
+    backstory="Expert in cloud security with 10 years of experience",
+    skills=["./skills/security-audit"],        # 도메인 지침
+    knowledge_sources=[pdf_source, url_source], # 검색된 사실
+)
+```
+
+---
+
+## 언제 무엇을 사용할까
+
+| 필요한 것...                                             | 사용할 것          | 예시                                     |
+| :------------------------------------------------------- | :---------------- | :--------------------------------------- |
+| 에이전트가 웹을 검색                                      | **도구**          | `tools=[SerperDevTool()]`                |
+| 에이전트가 MCP를 통해 원격 API 호출                       | **MCP**           | `mcps=["https://api.example.com/sse"]`   |
+| 에이전트가 Gmail로 이메일 전송                            | **앱**            | `apps=["gmail"]`                         |
+| 에이전트가 특정 절차를 따름                               | **스킬**          | `skills=["./skills/code-review"]`        |
+| 에이전트가 회사 문서 참조                                 | **지식**          | `knowledge_sources=[pdf_source]`         |
+| 에이전트가 웹 검색 AND 리뷰 가이드라인 준수               | **도구 + 스킬**    | 둘 다 함께 사용                           |
+
+---
+
+## 기능 조합하기
+
+실제로 에이전트는 종종 **여러 기능 유형을 함께** 사용합니다. 현실적인 예시입니다:
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+# 완전히 갖춘 리서치 에이전트
+researcher = Agent(
+    role="Senior Research Analyst",
+    goal="Produce comprehensive market analysis reports",
+    backstory="Expert analyst with deep industry knowledge",
+
+    # 액션: 에이전트가 할 수 있는 것
+    tools=[
+        SerperDevTool(),         # 웹 검색
+        FileReadTool(),          # 로컬 파일 읽기
+        CodeInterpreterTool(),   # 분석을 위한 Python 코드 실행
+    ],
+    mcps=["https://data-api.example.com/sse"],  # 원격 데이터 API 접근
+    apps=["google_sheets"],                      # Google Sheets에 쓰기
+
+    # 컨텍스트: 에이전트가 아는 것
+    skills=["./skills/research-methodology"],    # 연구 수행 방법
+    knowledge_sources=[company_docs],            # 회사 특화 데이터
+)
+```
+
+---
+
+## 비교 테이블
+
+| 특성 | 도구 | MCP | 앱 | 스킬 | 지식 |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **에이전트에게 액션 부여** | ✅ | ✅ | ✅ | ❌ | ❌ |
+| **프롬프트 수정** | ❌ | ❌ | ❌ | ✅ | ✅ |
+| **코드 필요** | 예 | 설정만 | 설정만 | 마크다운만 | 설정만 |
+| **로컬 실행** | 예 | 경우에 따라 | 예 (환경 변수 필요) | N/A | 예 |
+| **API 키 필요** | 도구별 | 서버별 | 통합 토큰 | 아니오 | 임베더만 |
+| **Agent에 설정** | `tools=[]` | `mcps=[]` | `apps=[]` | `skills=[]` | `knowledge_sources=[]` |
+| **Crew에 설정** | ❌ | ❌ | ❌ | `skills=[]` | `knowledge_sources=[]` |
+
+---
+
+## 상세 가이드
+
+각 기능 유형에 대해 더 알아볼 준비가 되셨나요?
+
+<CardGroup cols={2}>
+  <Card title="도구" icon="wrench" href="/ko/concepts/tools">
+    맞춤형 도구 생성, 75개 이상의 OSS 카탈로그 사용, 캐싱 및 비동기 실행 설정.
+  </Card>
+  <Card title="MCP 통합" icon="plug" href="/ko/mcp/overview">
+    stdio, SSE 또는 HTTP를 통해 MCP 서버에 연결. 도구 필터링, 인증 설정.
+  </Card>
+  <Card title="스킬" icon="bolt" href="/ko/concepts/skills">
+    SKILL.md로 스킬 패키지 구축, 도메인 전문성 주입, 점진적 공개 사용.
+  </Card>
+  <Card title="지식" icon="book" href="/ko/concepts/knowledge">
+    PDF, CSV, URL 등에서 지식 추가. 임베더 및 검색 설정.
+  </Card>
+</CardGroup>

docs/ko/concepts/skills.mdx

@@ -1,27 +1,186 @@
 ---
 title: 스킬
-description: 에이전트 프롬프트에 컨텍스트를 주입하는 파일 시스템 기반 스킬 패키지.
+description: 에이전트 프롬프트에 도메인 전문성과 지침을 주입하는 파일 시스템 기반 스킬 패키지.
 icon: bolt
 mode: "wide"
 ---
 
 ## 개요
 
-스킬은 에이전트에게 도메인별 지침, 참조 자료, 에셋을 제공하는 자체 포함 디렉터리입니다. 각 스킬은 YAML 프론트매터와 마크다운 본문이 포함된 `SKILL.md` 파일로 정의됩니다.
+스킬은 에이전트에게 **도메인별 지침, 가이드라인 및 참조 자료**를 제공하는 자체 포함 디렉터리입니다. 각 스킬은 YAML 프론트매터와 마크다운 본문이 포함된 `SKILL.md` 파일로 정의됩니다.
 
-스킬은 **점진적 공개**를 사용합니다 — 메타데이터가 먼저 로드되고, 활성화 시에만 전체 지침이 로드되며, 필요할 때만 리소스 카탈로그가 로드됩니다.
+활성화되면 스킬의 지침이 에이전트의 작업 프롬프트에 직접 주입됩니다 — 코드 변경 없이 에이전트에게 전문성을 부여합니다.
 
-## 디렉터리 구조
+<Note type="info" title="스킬 vs 도구 — 핵심 구분">
+**스킬은 도구가 아닙니다.** 이것이 가장 흔한 혼동 포인트입니다.
+
+- **스킬**은 에이전트의 프롬프트에 *지침과 컨텍스트*를 주입합니다. 에이전트에게 문제에 대해 *어떻게 생각할지*를 알려줍니다.
+- **도구**는 에이전트에게 행동을 취할 수 있는 *호출 가능한 함수*를 제공합니다 (검색, 파일 읽기, API 호출).
+
+흔히 **둘 다** 필요합니다: 전문성을 위한 스킬과 행동을 위한 도구. 이들은 독립적으로 구성되며 서로 보완합니다.
+</Note>
+
+---
+
+## 빠른 시작
+
+### 1. 스킬 디렉터리 생성
 
 ```
-my-skill/
-├── SKILL.md            # 필수 — 프론트매터 + 지침
-├── scripts/            # 선택 — 실행 가능한 스크립트
-├── references/         # 선택 — 참조 문서
-└── assets/             # 선택 — 정적 파일 (설정, 데이터)
+skills/
+└── code-review/
+    ├── SKILL.md          # 필수 — 지침
+    ├── references/       # 선택 — 참조 문서
+    │   └── style-guide.md
+    └── scripts/          # 선택 — 실행 가능한 스크립트
 ```
 
-디렉터리 이름은 `SKILL.md`의 `name` 필드와 일치해야 합니다.
+### 2. SKILL.md 작성
+
+```markdown
+---
+name: code-review
+description: Guidelines for conducting thorough code reviews with focus on security and performance.
+metadata:
+  author: your-team
+  version: "1.0"
+---
+
+## 코드 리뷰 가이드라인
+
+코드를 리뷰할 때 이 체크리스트를 따르세요:
+
+1. **보안**: 인젝션 취약점, 인증 우회, 데이터 노출 확인
+2. **성능**: N+1 쿼리, 불필요한 할당, 블로킹 호출 확인
+3. **가독성**: 명확한 네이밍, 적절한 주석, 일관된 스타일 보장
+4. **테스트**: 새로운 기능에 대한 적절한 테스트 커버리지 확인
+
+### 심각도 수준
+- **크리티컬**: 보안 취약점, 데이터 손실 위험 → 머지 차단
+- **메이저**: 성능 문제, 로직 오류 → 변경 요청
+- **마이너**: 스타일 문제, 네이밍 제안 → 코멘트와 함께 승인
+```
+
+### 3. 에이전트에 연결
+
+```python
+from crewai import Agent
+from crewai_tools import GithubSearchTool, FileReadTool
+
+reviewer = Agent(
+    role="Senior Code Reviewer",
+    goal="Review pull requests for quality and security issues",
+    backstory="Staff engineer with expertise in secure coding practices.",
+    skills=["./skills"],                          # 리뷰 가이드라인 주입
+    tools=[GithubSearchTool(), FileReadTool()],   # 에이전트가 코드를 읽을 수 있게 함
+)
+```
+
+이제 에이전트는 **전문성** (스킬에서)과 **기능** (도구에서) 모두를 갖추게 됩니다.
+
+---
+
+## 스킬 + 도구: 함께 작동하기
+
+스킬과 도구가 어떻게 보완하는지 보여주는 일반적인 패턴입니다:
+
+### 패턴 1: 스킬만 (도메인 전문성, 액션 불필요)
+
+에이전트가 특정 지침이 필요하지만 외부 서비스를 호출할 필요가 없을 때 사용:
+
+```python
+agent = Agent(
+    role="Technical Writer",
+    goal="Write clear API documentation",
+    backstory="Expert technical writer",
+    skills=["./skills/api-docs-style"],  # 작성 가이드라인 및 템플릿
+    # 도구 불필요 — 에이전트가 제공된 컨텍스트를 기반으로 작성
+)
+```
+
+### 패턴 2: 도구만 (액션, 특별한 전문성 불필요)
+
+에이전트가 행동을 취해야 하지만 도메인별 지침이 필요 없을 때 사용:
+
+```python
+from crewai_tools import SerperDevTool, ScrapeWebsiteTool
+
+agent = Agent(
+    role="Web Researcher",
+    goal="Find information about a topic",
+    backstory="Skilled at finding information online",
+    tools=[SerperDevTool(), ScrapeWebsiteTool()],  # 검색 및 스크래핑 가능
+    # 스킬 불필요 — 일반 연구에는 특별한 가이드라인이 필요 없음
+)
+```
+
+### 패턴 3: 스킬 + 도구 (전문성 AND 액션)
+
+가장 일반적인 실제 패턴. 스킬은 작업에 *어떻게* 접근할지를 제공하고, 도구는 에이전트가 *무엇을* 할 수 있는지를 제공합니다:
+
+```python
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+analyst = Agent(
+    role="Security Analyst",
+    goal="Audit infrastructure for vulnerabilities",
+    backstory="Expert in cloud security and compliance",
+    skills=["./skills/security-audit"],   # 감사 방법론 및 체크리스트
+    tools=[
+        SerperDevTool(),                  # 알려진 취약점 조사
+        FileReadTool(),                   # 설정 파일 읽기
+        CodeInterpreterTool(),            # 분석 스크립트 실행
+    ],
+)
+```
+
+### 패턴 4: 스킬 + MCP
+
+스킬은 도구와 마찬가지로 MCP 서버와 함께 작동합니다:
+
+```python
+agent = Agent(
+    role="Data Analyst",
+    goal="Analyze customer data and generate reports",
+    backstory="Expert data analyst with strong statistical background",
+    skills=["./skills/data-analysis"],           # 분석 방법론
+    mcps=["https://data-warehouse.example.com/sse"],  # 원격 데이터 접근
+)
+```
+
+### 패턴 5: 스킬 + 앱
+
+스킬은 에이전트가 플랫폼 통합을 사용하는 방법을 안내할 수 있습니다:
+
+```python
+agent = Agent(
+    role="Customer Support Agent",
+    goal="Respond to customer inquiries professionally",
+    backstory="Experienced support representative",
+    skills=["./skills/support-playbook"],  # 응답 템플릿 및 에스컬레이션 규칙
+    apps=["gmail", "zendesk"],             # 이메일 전송 및 티켓 업데이트 가능
+)
+```
+
+---
+
+## 크루 레벨 스킬
+
+스킬을 크루에 설정하여 **모든 에이전트**에 적용할 수 있습니다:
+
+```python
+from crewai import Crew
+
+crew = Crew(
+    agents=[researcher, writer, reviewer],
+    tasks=[research_task, write_task, review_task],
+    skills=["./skills"],  # 모든 에이전트가 이 스킬을 받음
+)
+```
+
+에이전트 레벨 스킬이 우선합니다 — 동일한 스킬이 양쪽 레벨에서 발견되면 에이전트의 버전이 사용됩니다.
+
+---
 
 ## SKILL.md 형식
 
@@ -34,7 +193,7 @@ compatibility: crewai>=0.1.0          # 선택
 metadata:                              # 선택
   author: your-name
   version: "1.0"
-allowed-tools: web-search file-read   # 선택, 공백으로 구분
+allowed-tools: web-search file-read   # 선택, 실험적
 ---
 
 에이전트를 위한 지침이 여기에 들어갑니다. 이 마크다운 본문은
@@ -43,57 +202,46 @@ allowed-tools: web-search file-read   # 선택, 공백으로 구분
 
 ### 프론트매터 필드
 
-| 필드             | 필수   | 제약 조건                                                                |
+| 필드             | 필수   | 설명                                                                     |
 | :-------------- | :----- | :----------------------------------------------------------------------- |
-| `name`          | 예     | 1–64자. 소문자 영숫자와 하이픈. 선행/후행/연속 하이픈 불가. 디렉터리 이름과 일치 필수. |
+| `name`          | 예     | 1–64자. 소문자 영숫자와 하이픈. 디렉터리 이름과 일치 필수.                  |
 | `description`   | 예     | 1–1024자. 스킬이 무엇을 하고 언제 사용하는지 설명.                          |
 | `license`       | 아니오  | 라이선스 이름 또는 번들된 라이선스 파일 참조.                                |
 | `compatibility` | 아니오  | 최대 500자. 환경 요구 사항 (제품, 패키지, 네트워크).                         |
 | `metadata`      | 아니오  | 임의의 문자열 키-값 매핑.                                                  |
 | `allowed-tools` | 아니오  | 공백으로 구분된 사전 승인 도구 목록. 실험적.                                 |
 
-## 사용법
+---
 
-### 에이전트 레벨 스킬
+## 디렉터리 구조
 
-에이전트에 스킬 디렉터리 경로를 전달합니다:
-
-```python
-from crewai import Agent
-
-agent = Agent(
-    role="Researcher",
-    goal="Find relevant information",
-    backstory="An expert researcher.",
-    skills=["./skills"],  # 이 디렉터리의 모든 스킬을 검색
-)
+```
+my-skill/
+├── SKILL.md            # 필수 — 프론트매터 + 지침
+├── scripts/            # 선택 — 실행 가능한 스크립트
+├── references/         # 선택 — 참조 문서
+└── assets/             # 선택 — 정적 파일 (설정, 데이터)
 ```
 
-### 크루 레벨 스킬
+디렉터리 이름은 `SKILL.md`의 `name` 필드와 일치해야 합니다. `scripts/`, `references/`, `assets/` 디렉터리는 파일을 직접 참조해야 하는 에이전트를 위해 스킬의 `path`에서 사용할 수 있습니다.
 
-크루의 스킬 경로는 모든 에이전트에 병합됩니다:
+---
 
-```python
-from crewai import Crew
+## 사전 로드된 스킬
 
-crew = Crew(
-    agents=[agent],
-    tasks=[task],
-    skills=["./skills"],
-)
-```
-
-### 사전 로드된 스킬
-
-`Skill` 객체를 직접 전달할 수도 있습니다:
+더 세밀한 제어를 위해 프로그래밍 방식으로 스킬을 검색하고 활성화할 수 있습니다:
 
 ```python
 from pathlib import Path
 from crewai.skills import discover_skills, activate_skill
 
+# 디렉터리의 모든 스킬 검색
 skills = discover_skills(Path("./skills"))
+
+# 활성화 (전체 SKILL.md 본문 로드)
 activated = [activate_skill(s) for s in skills]
 
+# 에이전트에 전달
 agent = Agent(
     role="Researcher",
     goal="Find relevant information",
@@ -102,13 +250,57 @@ agent = Agent(
 )
 ```
 
+---
+
 ## 스킬 로드 방식
 
-스킬은 점진적으로 로드됩니다 — 각 단계에서 필요한 데이터만 읽습니다:
+스킬은 **점진적 공개**를 사용합니다 — 각 단계에서 필요한 것만 로드합니다:
 
-| 단계              | 로드되는 내용                                      | 시점                |
-| :--------------- | :------------------------------------------------ | :----------------- |
-| 검색              | 이름, 설명, 프론트매터 필드                          | `discover_skills()` |
-| 활성화            | 전체 SKILL.md 본문 텍스트                           | `activate_skill()` |
+| 단계     | 로드되는 내용                          | 시점                |
+| :------- | :------------------------------------ | :------------------ |
+| 검색     | 이름, 설명, 프론트매터 필드             | `discover_skills()` |
+| 활성화   | 전체 SKILL.md 본문 텍스트              | `activate_skill()`  |
 
-일반적인 에이전트 실행 중에 스킬은 자동으로 검색되고 활성화됩니다. `scripts/`, `references/`, `assets/` 디렉터리는 파일을 직접 참조해야 하는 에이전트를 위해 스킬의 `path`에서 사용할 수 있습니다.
+일반적인 에이전트 실행 중(`skills=["./skills"]`로 디렉터리 경로 전달 시) 스킬은 자동으로 검색되고 활성화됩니다. 점진적 로딩은 프로그래밍 API를 사용할 때만 관련됩니다.
+
+---
+
+## 스킬 vs 지식
+
+스킬과 지식 모두 에이전트의 프롬프트를 수정하지만, 서로 다른 목적을 가지고 있습니다:
+
+| 측면 | 스킬 | 지식 |
+| :--- | :--- | :--- |
+| **제공하는 것** | 지침, 절차, 가이드라인 | 사실, 데이터, 정보 |
+| **저장 방식** | 마크다운 파일 (SKILL.md) | 벡터 스토어에 임베딩 (ChromaDB) |
+| **검색 방식** | 전체 본문이 프롬프트에 주입 | 시맨틱 검색으로 관련 청크 찾기 |
+| **적합한 용도** | 방법론, 체크리스트, 스타일 가이드 | 회사 문서, 제품 정보, 참조 데이터 |
+| **설정 방법** | `skills=["./skills"]` | `knowledge_sources=[source]` |
+
+**경험 법칙:** 에이전트가 *프로세스*를 따라야 하면 스킬을 사용하세요. 에이전트가 *데이터*를 참조해야 하면 지식을 사용하세요.
+
+---
+
+## 자주 묻는 질문
+
+<AccordionGroup>
+  <Accordion title="스킬과 도구를 모두 설정해야 하나요?">
+    사용 사례에 따라 다릅니다. 스킬과 도구는 **독립적**입니다 — 둘 중 하나, 둘 다, 또는 아무것도 사용하지 않을 수 있습니다.
+
+    - **스킬만**: 에이전트가 전문성은 필요하지만 외부 액션이 필요 없을 때 (예: 스타일 가이드라인으로 작성)
+    - **도구만**: 에이전트가 액션은 필요하지만 특별한 방법론이 필요 없을 때 (예: 간단한 웹 검색)
+    - **둘 다**: 에이전트가 전문성 AND 액션이 필요할 때 (예: 특정 체크리스트로 보안 감사 AND 코드 스캔 기능)
+  </Accordion>
+
+  <Accordion title="스킬이 자동으로 도구를 제공하나요?">
+    **아니요.** SKILL.md의 `allowed-tools` 필드는 실험적 메타데이터일 뿐 — 도구를 프로비저닝하거나 주입하지 않습니다. 항상 `tools=[]`, `mcps=[]` 또는 `apps=[]`를 통해 별도로 도구를 설정해야 합니다.
+  </Accordion>
+
+  <Accordion title="에이전트와 크루 모두에 같은 스킬을 설정하면 어떻게 되나요?">
+    에이전트 레벨 스킬이 우선합니다. 스킬은 이름으로 중복 제거됩니다 — 에이전트의 스킬이 먼저 처리되므로, 같은 스킬 이름이 양쪽 레벨에 나타나면 에이전트의 버전이 사용됩니다.
+  </Accordion>
+
+  <Accordion title="SKILL.md 본문의 최대 크기는 얼마인가요?">
+    50,000자에서 소프트 경고가 있지만 하드 리밋은 없습니다. 최상의 결과를 위해 스킬을 집중적이고 간결하게 유지하세요 — 너무 큰 프롬프트 주입은 에이전트의 주의를 분산시킬 수 있습니다.
+  </Accordion>
+</AccordionGroup>

docs/ko/concepts/tools.mdx

@@ -10,6 +10,10 @@ mode: "wide"
 CrewAI 도구는 에이전트에게 웹 검색, 데이터 분석부터 동료 간 협업 및 작업 위임에 이르기까지 다양한 기능을 제공합니다.  
 이 문서에서는 CrewAI 프레임워크 내에서 이러한 도구를 생성, 통합 및 활용하는 방법과, 협업 도구에 초점을 맞춘 새로운 기능에 대해 설명합니다.
 
+<Note type="info" title="도구는 다섯 가지 에이전트 기능 유형 중 하나입니다">
+  도구는 에이전트에게 행동을 취할 수 있는 **호출 가능한 함수**를 제공합니다. [MCP](/ko/mcp/overview) (원격 도구 서버), [앱](/ko/concepts/agent-capabilities) (플랫폼 통합), [스킬](/ko/concepts/skills) (도메인 전문성), [지식](/ko/concepts/knowledge) (검색된 사실)과 함께 작동합니다. 각 유형을 언제 사용해야 하는지 알아보려면 [에이전트 기능](/ko/concepts/agent-capabilities) 개요를 참조하세요.
+</Note>
+
 ## Tool이란 무엇인가?
 
 CrewAI에서 tool은 에이전트가 다양한 작업을 수행하기 위해 활용할 수 있는 기술 또는 기능입니다.

docs/ko/mcp/security.mdx

@@ -156,19 +156,7 @@ CrewAI 에이전트가 연결할 수 있는 MCP 서버를 개발하고 있다면
 - **MCP 인증 사양 준수**: 인증 및 권한 부여를 구현할 경우, [MCP Authorization specification](https://modelcontextprotocol.io/specification/draft/basic/authorization) 및 관련 [OAuth 2.0 security best practices](https://datatracker.ietf.org/doc/html/rfc9700)를 엄격히 준수하세요.
 - **정기적인 보안 감사**: MCP 서버가 민감한 데이터를 처리하거나, 중요한 작업을 수행하거나, 대외적으로 노출된 경우 자격을 갖춘 전문가의 정기적인 보안 감사를 고려하세요.
 
-## 5. 보안 취약점 보고
-
-CrewAI에서 보안 취약점을 발견하셨다면, Bugcrowd 취약점 공개 프로그램(VDP)을 통해 책임감 있게 보고해 주세요:
-
-**보고서 제출:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-공개 GitHub 이슈, 풀 리퀘스트 또는 소셜 미디어를 통해 취약점을 공개하지 **마세요**. Bugcrowd 이외의 채널로 제출된 보고서는 검토되지 않습니다.
-</Warning>
-
-자세한 내용은 [보안 정책](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md)을 참조하세요.
-
-## 6. 추가 참고 자료
+## 5. 추가 참고 자료
 
 MCP 보안에 대한 자세한 내용은 공식 문서를 참고하세요:
 - **[MCP 전송 보안](https://modelcontextprotocol.io/docs/concepts/transports#security-considerations)**

docs/ko/security.mdx

@@ -1,22 +0,0 @@
----
-title: 보안 정책
-description: CrewAI의 보안 취약점 보고 방법과 보안 관행에 대해 알아보세요.
-icon: shield
-mode: "wide"
----
-
-## 보안 취약점 보고
-
-CrewAI에서 보안 취약점을 발견하셨다면, Bugcrowd 취약점 공개 프로그램(VDP)을 통해 책임감 있게 보고해 주세요:
-
-**보고서 제출:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-공개 GitHub 이슈, 풀 리퀘스트 또는 소셜 미디어를 통해 취약점을 공개하지 **마세요**. Bugcrowd 이외의 채널로 제출된 보고서는 검토되지 않습니다.
-</Warning>
-
-자세한 내용은 [GitHub 보안 정책](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md)을 참조하세요.
-
-## 보안 리소스
-
-- **[MCP 보안 고려사항](/mcp/security)** — MCP 서버를 CrewAI 에이전트와 안전하게 통합하기 위한 모범 사례로, 전송 보안, 프롬프트 인젝션 위험 및 서버 구현 권장 사항을 포함합니다.

docs/pt-BR/changelog.mdx

@@ -4,6 +4,127 @@ description: "Atualizações de produto, melhorias e correções do CrewAI"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="31 mar 2026">
+  ## v1.13.0a5
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a5)
+
+  ## O que Mudou
+
+  ### Documentação
+  - Atualizar changelog e versão para v1.13.0a4
+
+  ## Contributors
+
+  @greysonlalonde, @joaomdmoura
+
+</Update>
+
+<Update label="01 abr 2026">
+  ## v1.13.0a4
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a4)
+
+  ## O que Mudou
+
+  ### Documentação
+  - Atualizar changelog e versão para v1.13.0a3
+
+  ## Contribuidores
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="01 abr 2026">
+  ## v1.13.0a3
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a3)
+
+  ## O que Mudou
+
+  ### Recursos
+  - Emitir dados de uso de token no LLMCallCompletedEvent
+  - Extrair e publicar metadados de ferramentas no AMP
+
+  ### Correções de Bugs
+  - Lidar com modelos GPT-5.x que não suportam o parâmetro de API `stop`
+
+  ### Documentação
+  - Corrigir imprecisões nas capacidades do agente em todas as línguas
+  - Adicionar visão geral das Capacidades do Agente e melhorar a documentação de Habilidades
+  - Adicionar um guia abrangente de configuração de SSO
+  - Atualizar o changelog e a versão para v1.13.0rc1
+
+  ### Refatoração
+  - Converter Flow para Pydantic BaseModel
+  - Converter classes LLM para Pydantic BaseModel
+  - Substituir InstanceOf[T] por anotações de tipo simples
+  - Remover métodos não utilizados
+
+  ## Contribuidores
+
+  @dependabot[bot], @greysonlalonde, @iris-clawd, @lorenzejay, @lucasgomide, @thiagomoretto
+
+</Update>
+
+<Update label="27 mar 2026">
+  ## v1.13.0rc1
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0rc1)
+
+  ## O que Mudou
+
+  ### Documentação
+  - Atualizar changelog e versão para v1.13.0a2
+
+  ## Contribuidores
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="27 mar 2026">
+  ## v1.13.0a2
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a2)
+
+  ## O que Mudou
+
+  ### Recursos
+  - Repositório de teste de implantação de autoatualização durante o lançamento
+  - Melhorar a resiliência e a experiência do usuário na versão empresarial
+
+  ### Documentação
+  - Atualizar changelog e versão para v1.13.0a1
+
+  ## Contribuidores
+
+  @greysonlalonde
+
+</Update>
+
+<Update label="27 mar 2026">
+  ## v1.13.0a1
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.13.0a1)
+
+  ## O que Mudou
+
+  ### Correções de Bugs
+  - Corrigir links quebrados no fluxo de documentação fixando o Node na LTS 22
+  - Limpar o cache uv para pacotes recém-publicados na versão empresarial
+
+  ### Documentação
+  - Adicionar uma matriz abrangente de permissões RBAC e guia de implantação
+  - Atualizar o changelog e a versão para v1.12.2
+
+  ## Contributors
+
+  @greysonlalonde, @iris-clawd, @joaomdmoura
+
+</Update>
+
 <Update label="25 mar 2026">
   ## v1.12.2
 

docs/pt-BR/concepts/agent-capabilities.mdx

@@ -0,0 +1,147 @@
+---
+title: "Capacidades do Agente"
+description: "Entenda as cinco formas de estender agentes CrewAI: Ferramentas, MCPs, Apps, Skills e Knowledge."
+icon: puzzle-piece
+mode: "wide"
+---
+
+## Visão Geral
+
+Agentes CrewAI podem ser estendidos com **cinco tipos distintos de capacidades**, cada um servindo a um propósito diferente. Entender quando usar cada um — e como eles funcionam juntos — é fundamental para construir agentes eficazes.
+
+<CardGroup cols={2}>
+  <Card title="Ferramentas" icon="wrench" href="/pt-BR/concepts/tools" color="#3B82F6">
+    **Funções chamáveis** — permitem que agentes tomem ações. Buscas na web, operações com arquivos, chamadas de API, execução de código.
+  </Card>
+  <Card title="Servidores MCP" icon="plug" href="/pt-BR/mcp/overview" color="#8B5CF6">
+    **Servidores de ferramentas remotos** — conectam agentes a servidores de ferramentas externos via Model Context Protocol. Mesmo efeito de ferramentas, mas hospedados externamente.
+  </Card>
+  <Card title="Apps" icon="grid-2" color="#EC4899">
+    **Integrações com plataformas** — conectam agentes a aplicativos SaaS (Gmail, Slack, Jira, Salesforce) via plataforma CrewAI. Executa localmente com um token de integração.
+  </Card>
+  <Card title="Skills" icon="bolt" href="/pt-BR/concepts/skills" color="#F59E0B">
+    **Expertise de domínio** — injetam instruções, diretrizes e material de referência nos prompts dos agentes. Skills dizem aos agentes *como pensar*.
+  </Card>
+  <Card title="Knowledge" icon="book" href="/pt-BR/concepts/knowledge" color="#10B981">
+    **Fatos recuperados** — fornecem aos agentes dados de documentos, arquivos e URLs via busca semântica (RAG). Knowledge dá aos agentes *o que saber*.
+  </Card>
+</CardGroup>
+
+---
+
+## A Distinção Fundamental
+
+O mais importante a entender: **essas capacidades se dividem em duas categorias**.
+
+### Capacidades de Ação (Ferramentas, MCPs, Apps)
+
+Estas dão aos agentes a capacidade de **fazer coisas** — chamar APIs, ler arquivos, buscar na web, enviar emails. No momento da execução, os três tipos se resolvem no mesmo formato interno (instâncias de `BaseTool`) e aparecem em uma lista unificada de ferramentas que o agente pode chamar.
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool
+
+agent = Agent(
+    role="Researcher",
+    goal="Find and compile market data",
+    backstory="Expert market analyst",
+    tools=[SerperDevTool(), FileReadTool()],  # Ferramentas locais
+    mcps=["https://mcp.example.com/sse"],     # Ferramentas de servidor MCP remoto
+    apps=["gmail", "google_sheets"],           # Integrações com plataformas
+)
+```
+
+### Capacidades de Contexto (Skills, Knowledge)
+
+Estas modificam o **prompt** do agente — injetando expertise, instruções ou dados recuperados antes do agente começar a raciocinar. Não dão aos agentes novas ações; elas moldam como os agentes pensam e a quais informações têm acesso.
+
+```python
+from crewai import Agent
+
+agent = Agent(
+    role="Security Auditor",
+    goal="Audit cloud infrastructure for vulnerabilities",
+    backstory="Expert in cloud security with 10 years of experience",
+    skills=["./skills/security-audit"],        # Instruções de domínio
+    knowledge_sources=[pdf_source, url_source], # Fatos recuperados
+)
+```
+
+---
+
+## Quando Usar o Quê
+
+| Você precisa...                                          | Use               | Exemplo                                  |
+| :------------------------------------------------------- | :---------------- | :--------------------------------------- |
+| Agente buscar na web                                     | **Ferramentas**   | `tools=[SerperDevTool()]`                |
+| Agente chamar uma API remota via MCP                     | **MCPs**          | `mcps=["https://api.example.com/sse"]`   |
+| Agente enviar emails pelo Gmail                          | **Apps**          | `apps=["gmail"]`                         |
+| Agente seguir procedimentos específicos                  | **Skills**        | `skills=["./skills/code-review"]`        |
+| Agente consultar documentos da empresa                   | **Knowledge**     | `knowledge_sources=[pdf_source]`         |
+| Agente buscar na web E seguir diretrizes de revisão      | **Ferramentas + Skills** | Use ambos juntos                  |
+
+---
+
+## Combinando Capacidades
+
+Na prática, agentes frequentemente usam **múltiplos tipos de capacidades juntos**. Aqui está um exemplo realista:
+
+```python
+from crewai import Agent
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+# Um agente de pesquisa totalmente equipado
+researcher = Agent(
+    role="Senior Research Analyst",
+    goal="Produce comprehensive market analysis reports",
+    backstory="Expert analyst with deep industry knowledge",
+
+    # AÇÃO: O que o agente pode FAZER
+    tools=[
+        SerperDevTool(),         # Buscar na web
+        FileReadTool(),          # Ler arquivos locais
+        CodeInterpreterTool(),   # Executar código Python para análise
+    ],
+    mcps=["https://data-api.example.com/sse"],  # Acessar API de dados remota
+    apps=["google_sheets"],                      # Escrever no Google Sheets
+
+    # CONTEXTO: O que o agente SABE
+    skills=["./skills/research-methodology"],    # Como conduzir pesquisas
+    knowledge_sources=[company_docs],            # Dados específicos da empresa
+)
+```
+
+---
+
+## Tabela Comparativa
+
+| Característica | Ferramentas | MCPs | Apps | Skills | Knowledge |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **Dá ações ao agente** | ✅ | ✅ | ✅ | ❌ | ❌ |
+| **Modifica o prompt** | ❌ | ❌ | ❌ | ✅ | ✅ |
+| **Requer código** | Sim | Apenas config | Apenas config | Apenas Markdown | Apenas config |
+| **Executa localmente** | Sim | Depende | Sim (com variável de ambiente) | N/A | Sim |
+| **Precisa de chaves API** | Por ferramenta | Por servidor | Token de integração | Não | Apenas embedder |
+| **Definido no Agent** | `tools=[]` | `mcps=[]` | `apps=[]` | `skills=[]` | `knowledge_sources=[]` |
+| **Definido no Crew** | ❌ | ❌ | ❌ | `skills=[]` | `knowledge_sources=[]` |
+
+---
+
+## Aprofundamentos
+
+Pronto para aprender mais sobre cada tipo de capacidade?
+
+<CardGroup cols={2}>
+  <Card title="Ferramentas" icon="wrench" href="/pt-BR/concepts/tools">
+    Crie ferramentas personalizadas, use o catálogo OSS com 75+ opções, configure cache e execução assíncrona.
+  </Card>
+  <Card title="Integração MCP" icon="plug" href="/pt-BR/mcp/overview">
+    Conecte-se a servidores MCP via stdio, SSE ou HTTP. Filtre ferramentas, configure autenticação.
+  </Card>
+  <Card title="Skills" icon="bolt" href="/pt-BR/concepts/skills">
+    Construa pacotes de skills com SKILL.md, injete expertise de domínio, use divulgação progressiva.
+  </Card>
+  <Card title="Knowledge" icon="book" href="/pt-BR/concepts/knowledge">
+    Adicione conhecimento de PDFs, CSVs, URLs e mais. Configure embedders e recuperação.
+  </Card>
+</CardGroup>

docs/pt-BR/concepts/skills.mdx

@@ -1,27 +1,186 @@
 ---
 title: Skills
-description: Pacotes de skills baseados em sistema de arquivos que injetam contexto nos prompts dos agentes.
+description: Pacotes de skills baseados em sistema de arquivos que injetam expertise de domínio e instruções nos prompts dos agentes.
 icon: bolt
 mode: "wide"
 ---
 
 ## Visão Geral
 
-Skills são diretórios autocontidos que fornecem aos agentes instruções, referências e assets específicos de domínio. Cada skill é definida por um arquivo `SKILL.md` com frontmatter YAML e um corpo em markdown.
+Skills são diretórios autocontidos que fornecem aos agentes **instruções, diretrizes e material de referência específicos de domínio**. Cada skill é definida por um arquivo `SKILL.md` com frontmatter YAML e um corpo em markdown.
 
-Skills usam **divulgação progressiva** — metadados são carregados primeiro, instruções completas apenas quando ativadas, e catálogos de recursos apenas quando necessário.
+Quando ativada, as instruções de uma skill são injetadas diretamente no prompt da tarefa do agente — dando ao agente expertise sem exigir alterações de código.
 
-## Estrutura de Diretório
+<Note type="info" title="Skills vs Ferramentas — A Distinção Fundamental">
+**Skills NÃO são ferramentas.** Este é o ponto de confusão mais comum.
+
+- **Skills** injetam *instruções e contexto* no prompt do agente. Elas dizem ao agente *como pensar* sobre um problema.
+- **Ferramentas** dão ao agente *funções chamáveis* para tomar ações (buscar, ler arquivos, chamar APIs).
+
+Frequentemente você precisa de **ambos**: skills para expertise, ferramentas para ação. Eles são configurados independentemente e se complementam.
+</Note>
+
+---
+
+## Início Rápido
+
+### 1. Crie um Diretório de Skill
 
 ```
-my-skill/
-├── SKILL.md            # Obrigatório — frontmatter + instruções
-├── scripts/            # Opcional — scripts executáveis
-├── references/         # Opcional — documentos de referência
-└── assets/             # Opcional — arquivos estáticos (configs, dados)
+skills/
+└── code-review/
+    ├── SKILL.md          # Obrigatório — instruções
+    ├── references/       # Opcional — documentos de referência
+    │   └── style-guide.md
+    └── scripts/          # Opcional — scripts executáveis
 ```
 
-O nome do diretório deve corresponder ao campo `name` no `SKILL.md`.
+### 2. Escreva seu SKILL.md
+
+```markdown
+---
+name: code-review
+description: Guidelines for conducting thorough code reviews with focus on security and performance.
+metadata:
+  author: your-team
+  version: "1.0"
+---
+
+## Diretrizes de Code Review
+
+Ao revisar código, siga esta checklist:
+
+1. **Segurança**: Verifique vulnerabilidades de injeção, bypasses de autenticação e exposição de dados
+2. **Performance**: Procure por queries N+1, alocações desnecessárias e chamadas bloqueantes
+3. **Legibilidade**: Garanta nomenclatura clara, comentários apropriados e estilo consistente
+4. **Testes**: Verifique cobertura adequada de testes para novas funcionalidades
+
+### Níveis de Severidade
+- **Crítico**: Vulnerabilidades de segurança, riscos de perda de dados → bloquear merge
+- **Major**: Problemas de performance, erros de lógica → solicitar alterações
+- **Minor**: Questões de estilo, sugestões de nomenclatura → aprovar com comentários
+```
+
+### 3. Anexe a um Agente
+
+```python
+from crewai import Agent
+from crewai_tools import GithubSearchTool, FileReadTool
+
+reviewer = Agent(
+    role="Senior Code Reviewer",
+    goal="Review pull requests for quality and security issues",
+    backstory="Staff engineer with expertise in secure coding practices.",
+    skills=["./skills"],                          # Injeta diretrizes de revisão
+    tools=[GithubSearchTool(), FileReadTool()],   # Permite ao agente ler código
+)
+```
+
+O agente agora tem tanto **expertise** (da skill) quanto **capacidades** (das ferramentas).
+
+---
+
+## Skills + Ferramentas: Trabalhando Juntos
+
+Aqui estão padrões comuns mostrando como skills e ferramentas se complementam:
+
+### Padrão 1: Apenas Skills (Expertise de Domínio, Sem Ações Necessárias)
+
+Use quando o agente precisa de instruções específicas mas não precisa chamar serviços externos:
+
+```python
+agent = Agent(
+    role="Technical Writer",
+    goal="Write clear API documentation",
+    backstory="Expert technical writer",
+    skills=["./skills/api-docs-style"],  # Diretrizes e templates de escrita
+    # Sem ferramentas necessárias — agente escreve baseado no contexto fornecido
+)
+```
+
+### Padrão 2: Apenas Ferramentas (Ações, Sem Expertise Especial)
+
+Use quando o agente precisa tomar ações mas não precisa de instruções específicas de domínio:
+
+```python
+from crewai_tools import SerperDevTool, ScrapeWebsiteTool
+
+agent = Agent(
+    role="Web Researcher",
+    goal="Find information about a topic",
+    backstory="Skilled at finding information online",
+    tools=[SerperDevTool(), ScrapeWebsiteTool()],  # Pode buscar e extrair dados
+    # Sem skills necessárias — pesquisa geral não precisa de diretrizes especiais
+)
+```
+
+### Padrão 3: Skills + Ferramentas (Expertise E Ações)
+
+O padrão mais comum no mundo real. A skill fornece *como* abordar o trabalho; ferramentas fornecem *o que* o agente pode fazer:
+
+```python
+from crewai_tools import SerperDevTool, FileReadTool, CodeInterpreterTool
+
+analyst = Agent(
+    role="Security Analyst",
+    goal="Audit infrastructure for vulnerabilities",
+    backstory="Expert in cloud security and compliance",
+    skills=["./skills/security-audit"],   # Metodologia e checklists de auditoria
+    tools=[
+        SerperDevTool(),                  # Pesquisar vulnerabilidades conhecidas
+        FileReadTool(),                   # Ler arquivos de configuração
+        CodeInterpreterTool(),            # Executar scripts de análise
+    ],
+)
+```
+
+### Padrão 4: Skills + MCPs
+
+Skills funcionam junto com servidores MCP da mesma forma que com ferramentas:
+
+```python
+agent = Agent(
+    role="Data Analyst",
+    goal="Analyze customer data and generate reports",
+    backstory="Expert data analyst with strong statistical background",
+    skills=["./skills/data-analysis"],           # Metodologia de análise
+    mcps=["https://data-warehouse.example.com/sse"],  # Acesso remoto a dados
+)
+```
+
+### Padrão 5: Skills + Apps
+
+Skills podem guiar como um agente usa integrações de plataforma:
+
+```python
+agent = Agent(
+    role="Customer Support Agent",
+    goal="Respond to customer inquiries professionally",
+    backstory="Experienced support representative",
+    skills=["./skills/support-playbook"],  # Templates de resposta e regras de escalação
+    apps=["gmail", "zendesk"],             # Pode enviar emails e atualizar tickets
+)
+```
+
+---
+
+## Skills no Nível do Crew
+
+Skills podem ser definidas no crew para aplicar a **todos os agentes**:
+
+```python
+from crewai import Crew
+
+crew = Crew(
+    agents=[researcher, writer, reviewer],
+    tasks=[research_task, write_task, review_task],
+    skills=["./skills"],  # Todos os agentes recebem essas skills
+)
+```
+
+Skills no nível do agente têm prioridade — se a mesma skill é descoberta em ambos os níveis, a versão do agente é usada.
+
+---
 
 ## Formato do SKILL.md
 
@@ -34,7 +193,7 @@ compatibility: crewai>=0.1.0          # opcional
 metadata:                              # opcional
   author: your-name
   version: "1.0"
-allowed-tools: web-search file-read   # opcional, delimitado por espaços
+allowed-tools: web-search file-read   # opcional, experimental
 ---
 
 Instruções para o agente vão aqui. Este corpo em markdown é injetado
@@ -43,57 +202,46 @@ no prompt do agente quando a skill é ativada.
 
 ### Campos do Frontmatter
 
-| Campo           | Obrigatório | Restrições                                                               |
+| Campo           | Obrigatório | Descrição                                                                |
 | :-------------- | :---------- | :----------------------------------------------------------------------- |
-| `name`          | Sim         | 1–64 chars. Alfanumérico minúsculo e hifens. Sem hifens iniciais/finais/consecutivos. Deve corresponder ao nome do diretório. |
+| `name`          | Sim         | 1–64 chars. Alfanumérico minúsculo e hifens. Deve corresponder ao nome do diretório. |
 | `description`   | Sim         | 1–1024 chars. Descreve o que a skill faz e quando usá-la.                |
 | `license`       | Não         | Nome da licença ou referência a um arquivo de licença incluído.           |
 | `compatibility` | Não         | Máx 500 chars. Requisitos de ambiente (produtos, pacotes, rede).          |
 | `metadata`      | Não         | Mapeamento arbitrário de chave-valor string.                              |
 | `allowed-tools` | Não         | Lista de ferramentas pré-aprovadas delimitada por espaços. Experimental.  |
 
-## Uso
+---
 
-### Skills no Nível do Agente
+## Estrutura de Diretório
 
-Passe caminhos de diretório de skills para um agente:
-
-```python
-from crewai import Agent
-
-agent = Agent(
-    role="Researcher",
-    goal="Find relevant information",
-    backstory="An expert researcher.",
-    skills=["./skills"],  # descobre todas as skills neste diretório
-)
+```
+my-skill/
+├── SKILL.md            # Obrigatório — frontmatter + instruções
+├── scripts/            # Opcional — scripts executáveis
+├── references/         # Opcional — documentos de referência
+└── assets/             # Opcional — arquivos estáticos (configs, dados)
 ```
 
-### Skills no Nível do Crew
+O nome do diretório deve corresponder ao campo `name` no `SKILL.md`. Os diretórios `scripts/`, `references/` e `assets/` estão disponíveis no `path` da skill para agentes que precisam referenciar arquivos diretamente.
 
-Caminhos de skills no crew são mesclados em todos os agentes:
+---
 
-```python
-from crewai import Crew
+## Skills Pré-carregadas
 
-crew = Crew(
-    agents=[agent],
-    tasks=[task],
-    skills=["./skills"],
-)
-```
-
-### Skills Pré-carregadas
-
-Você também pode passar objetos `Skill` diretamente:
+Para mais controle, você pode descobrir e ativar skills programaticamente:
 
 ```python
 from pathlib import Path
 from crewai.skills import discover_skills, activate_skill
 
+# Descobrir todas as skills em um diretório
 skills = discover_skills(Path("./skills"))
+
+# Ativá-las (carrega o corpo completo do SKILL.md)
 activated = [activate_skill(s) for s in skills]
 
+# Passar para um agente
 agent = Agent(
     role="Researcher",
     goal="Find relevant information",
@@ -102,13 +250,57 @@ agent = Agent(
 )
 ```
 
+---
+
 ## Como as Skills São Carregadas
 
-Skills carregam progressivamente — apenas os dados necessários em cada etapa são lidos:
+Skills usam **divulgação progressiva** — carregando apenas o necessário em cada estágio:
 
-| Etapa            | O que é carregado                                  | Quando              |
-| :--------------- | :------------------------------------------------ | :------------------ |
-| Descoberta       | Nome, descrição, campos do frontmatter             | `discover_skills()` |
-| Ativação         | Texto completo do corpo do SKILL.md                | `activate_skill()` |
+| Estágio    | O que é carregado                     | Quando              |
+| :--------- | :------------------------------------ | :------------------ |
+| Descoberta | Nome, descrição, campos do frontmatter | `discover_skills()` |
+| Ativação   | Texto completo do corpo do SKILL.md    | `activate_skill()`  |
 
-Durante a execução normal do agente, skills são automaticamente descobertas e ativadas. Os diretórios `scripts/`, `references/` e `assets/` estão disponíveis no `path` da skill para agentes que precisam referenciar arquivos diretamente.
+Durante a execução normal do agente (passando caminhos de diretório via `skills=["./skills"]`), skills são automaticamente descobertas e ativadas. O carregamento progressivo só importa quando usando a API programática.
+
+---
+
+## Skills vs Knowledge
+
+Tanto skills quanto knowledge modificam o prompt do agente, mas servem propósitos diferentes:
+
+| Aspecto | Skills | Knowledge |
+| :--- | :--- | :--- |
+| **O que fornece** | Instruções, procedimentos, diretrizes | Fatos, dados, informações |
+| **Como é armazenado** | Arquivos Markdown (SKILL.md) | Embarcado em banco vetorial (ChromaDB) |
+| **Como é recuperado** | Corpo inteiro injetado no prompt | Busca semântica encontra trechos relevantes |
+| **Melhor para** | Metodologia, checklists, guias de estilo | Documentos da empresa, info de produto, dados de referência |
+| **Definido via** | `skills=["./skills"]` | `knowledge_sources=[source]` |
+
+**Regra prática:** Se o agente precisa seguir um *processo*, use uma skill. Se o agente precisa consultar *dados*, use knowledge.
+
+---
+
+## Perguntas Frequentes
+
+<AccordionGroup>
+  <Accordion title="Preciso definir skills E ferramentas?">
+    Depende do seu caso de uso. Skills e ferramentas são **independentes** — você pode usar qualquer um, ambos ou nenhum.
+
+    - **Apenas skills**: Quando o agente precisa de expertise mas não de ações externas (ex: escrever com diretrizes de estilo)
+    - **Apenas ferramentas**: Quando o agente precisa de ações mas não de metodologia especial (ex: busca simples na web)
+    - **Ambos**: Quando o agente precisa de expertise E ações (ex: auditoria de segurança com checklists específicas E capacidade de escanear código)
+  </Accordion>
+
+  <Accordion title="Skills fornecem ferramentas automaticamente?">
+    **Não.** O campo `allowed-tools` no SKILL.md é apenas metadado experimental — ele não provisiona nem injeta nenhuma ferramenta. Você deve sempre definir ferramentas separadamente via `tools=[]`, `mcps=[]` ou `apps=[]`.
+  </Accordion>
+
+  <Accordion title="O que acontece se eu definir a mesma skill tanto no agente quanto no crew?">
+    A skill no nível do agente tem prioridade. Skills são deduplicadas por nome — as skills do agente são processadas primeiro, então se o mesmo nome de skill aparece em ambos os níveis, a versão do agente é usada.
+  </Accordion>
+
+  <Accordion title="Qual o tamanho máximo do corpo do SKILL.md?">
+    Há um aviso suave em 50.000 caracteres, mas sem limite rígido. Mantenha skills focadas e concisas para melhores resultados — injeções de prompt muito grandes podem diluir a atenção do agente.
+  </Accordion>
+</AccordionGroup>

docs/pt-BR/concepts/tools.mdx

@@ -10,6 +10,10 @@ mode: "wide"
 As ferramentas do CrewAI capacitam agentes com habilidades que vão desde busca na web e análise de dados até colaboração e delegação de tarefas entre colegas de trabalho.
 Esta documentação descreve como criar, integrar e aproveitar essas ferramentas dentro do framework CrewAI, incluindo um novo foco em ferramentas de colaboração.
 
+<Note type="info" title="Ferramentas são um dos cinco tipos de capacidades de agentes">
+  Ferramentas dão aos agentes **funções chamáveis** para tomar ações. Elas funcionam junto com [MCPs](/pt-BR/mcp/overview) (servidores de ferramentas remotos), [Apps](/pt-BR/concepts/agent-capabilities) (integrações com plataformas), [Skills](/pt-BR/concepts/skills) (expertise de domínio) e [Knowledge](/pt-BR/concepts/knowledge) (fatos recuperados). Veja a visão geral de [Capacidades do Agente](/pt-BR/concepts/agent-capabilities) para entender quando usar cada um.
+</Note>
+
 ## O que é uma Ferramenta?
 
 Uma ferramenta no CrewAI é uma habilidade ou função que os agentes podem utilizar para executar diversas ações.

docs/pt-BR/mcp/security.mdx

@@ -156,19 +156,7 @@ Se você está desenvolvendo um servidor MCP ao qual agentes CrewAI possam se co
 - **Aderência à Especificação de Autorização MCP**: Caso implemente autenticação e autorização, siga estritamente a [especificação de autorização MCP](https://modelcontextprotocol.io/specification/draft/basic/authorization) e as [melhores práticas de segurança OAuth 2.0](https://datatracker.ietf.org/doc/html/rfc9700) relevantes.
 - **Auditorias de Segurança Regulares**: Caso seu servidor MCP manipule dados sensíveis, realize operações críticas ou seja exposto publicamente, considere auditorias de segurança periódicas conduzidas por profissionais qualificados.
 
-## 5. Reportando Vulnerabilidades de Segurança
-
-Se você descobrir uma vulnerabilidade de segurança no CrewAI, por favor reporte de forma responsável através do nosso Programa de Divulgação de Vulnerabilidades (VDP) no Bugcrowd:
-
-**Envie relatórios para:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-**Não** divulgue vulnerabilidades por meio de issues públicas no GitHub, pull requests ou redes sociais. Relatórios enviados por outros canais que não o Bugcrowd não serão analisados.
-</Warning>
-
-Para mais detalhes, consulte nossa [Política de Segurança](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md).
-
-## 6. Leituras Adicionais
+## 5. Leituras Adicionais
 
 Para informações mais detalhadas sobre segurança MCP, consulte a documentação oficial:
 - **[Segurança de Transporte MCP](https://modelcontextprotocol.io/docs/concepts/transports#security-considerations)**

docs/pt-BR/security.mdx

@@ -1,22 +0,0 @@
----
-title: Política de Segurança
-description: Saiba como reportar vulnerabilidades de segurança e sobre as práticas de segurança do CrewAI.
-icon: shield
-mode: "wide"
----
-
-## Reportando Vulnerabilidades de Segurança
-
-Se você descobrir uma vulnerabilidade de segurança no CrewAI, por favor reporte de forma responsável através do nosso Programa de Divulgação de Vulnerabilidades (VDP) no Bugcrowd:
-
-**Envie relatórios para:** [crewai-vdp-ess@submit.bugcrowd.com](mailto:crewai-vdp-ess@submit.bugcrowd.com)
-
-<Warning>
-**Não** divulgue vulnerabilidades por meio de issues públicas no GitHub, pull requests ou redes sociais. Relatórios enviados por outros canais que não o Bugcrowd não serão analisados.
-</Warning>
-
-Para mais detalhes, consulte nossa [Política de Segurança no GitHub](https://github.com/crewAIInc/crewAI/blob/main/.github/security.md).
-
-## Recursos de Segurança
-
-- **[Considerações de Segurança MCP](/mcp/security)** — Melhores práticas para integrar servidores MCP com segurança aos seus agentes CrewAI, incluindo segurança de transporte, riscos de injeção de prompt e conselhos de implementação de servidor.

lib/crewai-files/src/crewai_files/__init__.py

@@ -152,4 +152,4 @@ __all__ = [
     "wrap_file_source",
 ]
 
-__version__ = "1.12.2"
+__version__ = "1.13.0a5"

lib/crewai-tools/pyproject.toml

@@ -11,7 +11,7 @@ dependencies = [
     "pytube~=15.0.0",
     "requests~=2.32.5",
     "docker~=7.1.0",
-    "crewai==1.12.2",
+    "crewai==1.13.0a5",
     "tiktoken~=0.8.0",
     "beautifulsoup4~=4.13.4",
     "python-docx~=1.2.0",

lib/crewai-tools/src/crewai_tools/__init__.py

@@ -309,4 +309,4 @@ __all__ = [
     "ZapierActionTools",
 ]
 
-__version__ = "1.12.2"
+__version__ = "1.13.0a5"

lib/crewai-tools/tool.specs.json

@@ -14281,10 +14281,349 @@
             ],
             "title": "EnvVar",
             "type": "object"
+          },
+          "JsonResponseFormat": {
+            "description": "Response format requesting raw JSON output (e.g. ``{\"type\": \"json_object\"}``).",
+            "properties": {
+              "type": {
+                "const": "json_object",
+                "title": "Type",
+                "type": "string"
+              }
+            },
+            "required": [
+              "type"
+            ],
+            "title": "JsonResponseFormat",
+            "type": "object"
+          },
+          "LLM": {
+            "properties": {
+              "additional_params": {
+                "additionalProperties": true,
+                "title": "Additional Params",
+                "type": "object"
+              },
+              "api_base": {
+                "anyOf": [
+                  {
+                    "type": "string"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Api Base"
+              },
+              "api_key": {
+                "anyOf": [
+                  {
+                    "type": "string"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Api Key"
+              },
+              "api_version": {
+                "anyOf": [
+                  {
+                    "type": "string"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Api Version"
+              },
+              "base_url": {
+                "anyOf": [
+                  {
+                    "type": "string"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Base Url"
+              },
+              "callbacks": {
+                "anyOf": [
+                  {
+                    "items": {},
+                    "type": "array"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Callbacks"
+              },
+              "completion_cost": {
+                "anyOf": [
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Completion Cost"
+              },
+              "context_window_size": {
+                "default": 0,
+                "title": "Context Window Size",
+                "type": "integer"
+              },
+              "frequency_penalty": {
+                "anyOf": [
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Frequency Penalty"
+              },
+              "interceptor": {
+                "default": null,
+                "title": "Interceptor"
+              },
+              "is_anthropic": {
+                "default": false,
+                "title": "Is Anthropic",
+                "type": "boolean"
+              },
+              "is_litellm": {
+                "default": false,
+                "title": "Is Litellm",
+                "type": "boolean"
+              },
+              "logit_bias": {
+                "anyOf": [
+                  {
+                    "additionalProperties": {
+                      "type": "number"
+                    },
+                    "type": "object"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Logit Bias"
+              },
+              "logprobs": {
+                "anyOf": [
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Logprobs"
+              },
+              "max_completion_tokens": {
+                "anyOf": [
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Max Completion Tokens"
+              },
+              "max_tokens": {
+                "anyOf": [
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Max Tokens"
+              },
+              "model": {
+                "title": "Model",
+                "type": "string"
+              },
+              "n": {
+                "anyOf": [
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "N"
+              },
+              "prefer_upload": {
+                "default": false,
+                "title": "Prefer Upload",
+                "type": "boolean"
+              },
+              "presence_penalty": {
+                "anyOf": [
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Presence Penalty"
+              },
+              "provider": {
+                "default": "openai",
+                "title": "Provider",
+                "type": "string"
+              },
+              "reasoning_effort": {
+                "anyOf": [
+                  {
+                    "enum": [
+                      "none",
+                      "low",
+                      "medium",
+                      "high"
+                    ],
+                    "type": "string"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Reasoning Effort"
+              },
+              "response_format": {
+                "anyOf": [
+                  {
+                    "$ref": "#/$defs/JsonResponseFormat"
+                  },
+                  {},
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Response Format"
+              },
+              "seed": {
+                "anyOf": [
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Seed"
+              },
+              "stop": {
+                "items": {
+                  "type": "string"
+                },
+                "title": "Stop",
+                "type": "array"
+              },
+              "stream": {
+                "default": false,
+                "title": "Stream",
+                "type": "boolean"
+              },
+              "temperature": {
+                "anyOf": [
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Temperature"
+              },
+              "thinking": {
+                "default": null,
+                "title": "Thinking"
+              },
+              "timeout": {
+                "anyOf": [
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Timeout"
+              },
+              "top_logprobs": {
+                "anyOf": [
+                  {
+                    "type": "integer"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Top Logprobs"
+              },
+              "top_p": {
+                "anyOf": [
+                  {
+                    "type": "number"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "default": null,
+                "title": "Top P"
+              }
+            },
+            "required": [
+              "model"
+            ],
+            "title": "LLM",
+            "type": "object"
           }
         },
         "description": "A tool for performing Optical Character Recognition on images.\n\nThis tool leverages LLMs to extract text from images. It can process\nboth local image files and images available via URLs.\n\nAttributes:\n    name (str): Name of the tool.\n    description (str): Description of the tool's functionality.\n    args_schema (Type[BaseModel]): Pydantic schema for input validation.\n\nPrivate Attributes:\n    _llm (Optional[LLM]): Language model instance for making API calls.",
-        "properties": {},
+        "properties": {
+          "llm": {
+            "$ref": "#/$defs/LLM"
+          }
+        },
         "title": "OCRTool",
         "type": "object"
       },

lib/crewai/pyproject.toml

@@ -43,7 +43,7 @@ dependencies = [
     "uv~=0.9.13",
     "aiosqlite~=0.21.0",
     "pyyaml~=6.0",
-    "lancedb>=0.29.2",
+    "lancedb>=0.29.2,<0.30.1",
 ]
 
 [project.urls]
@@ -54,7 +54,7 @@ Repository = "https://github.com/crewAIInc/crewAI"
 
 [project.optional-dependencies]
 tools = [
-    "crewai-tools==1.12.2",
+    "crewai-tools==1.13.0a5",
 ]
 embeddings = [
     "tiktoken~=0.8.0"

lib/crewai/src/crewai/__init__.py

@@ -4,6 +4,8 @@ from typing import Any
 import urllib.request
 import warnings
 
+from pydantic import PydanticUserError
+
 from crewai.agent.core import Agent
 from crewai.agent.planning_config import PlanningConfig
 from crewai.crew import Crew
@@ -42,7 +44,7 @@ def _suppress_pydantic_deprecation_warnings() -> None:
 
 _suppress_pydantic_deprecation_warnings()
 
-__version__ = "1.12.2"
+__version__ = "1.13.0a5"
 _telemetry_submitted = False
 
 
@@ -93,6 +95,38 @@ def __getattr__(name: str) -> Any:
     raise AttributeError(f"module 'crewai' has no attribute {name!r}")
 
 
+try:
+    from crewai.agents.tools_handler import ToolsHandler as _ToolsHandler
+    from crewai.experimental.agent_executor import AgentExecutor as _AgentExecutor
+    from crewai.hooks.llm_hooks import LLMCallHookContext as _LLMCallHookContext
+    from crewai.tools.tool_types import ToolResult as _ToolResult
+    from crewai.utilities.prompts import (
+        StandardPromptResult as _StandardPromptResult,
+        SystemPromptResult as _SystemPromptResult,
+    )
+
+    _AgentExecutor.model_rebuild(
+        force=True,
+        _types_namespace={
+            "Agent": Agent,
+            "ToolsHandler": _ToolsHandler,
+            "Crew": Crew,
+            "BaseLLM": BaseLLM,
+            "Task": Task,
+            "StandardPromptResult": _StandardPromptResult,
+            "SystemPromptResult": _SystemPromptResult,
+            "LLMCallHookContext": _LLMCallHookContext,
+            "ToolResult": _ToolResult,
+        },
+    )
+except (ImportError, PydanticUserError):
+    import logging as _logging
+
+    _logging.getLogger(__name__).warning(
+        "AgentExecutor.model_rebuild() failed; forward refs may be unresolved.",
+        exc_info=True,
+    )
+
 __all__ = [
     "LLM",
     "Agent",

lib/crewai/src/crewai/agent/core.py

@@ -25,7 +25,6 @@ from pydantic import (
     BaseModel,
     ConfigDict,
     Field,
-    InstanceOf,
     PrivateAttr,
     model_validator,
 )
@@ -167,10 +166,10 @@ class Agent(BaseAgent):
         default=True,
         description="Use system prompt for the agent.",
     )
-    llm: str | InstanceOf[BaseLLM] | None = Field(
+    llm: str | BaseLLM | None = Field(
         description="Language model that will run the agent.", default=None
     )
-    function_calling_llm: str | InstanceOf[BaseLLM] | None = Field(
+    function_calling_llm: str | BaseLLM | None = Field(
         description="Language model that will run the agent.", default=None
     )
     system_template: str | None = Field(
@@ -1012,7 +1011,7 @@ class Agent(BaseAgent):
         self.agent_executor.tools = tools
         self.agent_executor.original_tools = raw_tools
         self.agent_executor.prompt = prompt
-        self.agent_executor.stop = stop_words
+        self.agent_executor.stop_words = stop_words
         self.agent_executor.tools_names = get_tool_names(tools)
         self.agent_executor.tools_description = render_text_description_and_args(tools)
         self.agent_executor.response_model = (

lib/crewai/src/crewai/agents/agent_builder/base_agent.py

@@ -12,7 +12,6 @@ from pydantic import (
     UUID4,
     BaseModel,
     Field,
-    InstanceOf,
     PrivateAttr,
     field_validator,
     model_validator,
@@ -185,7 +184,7 @@ class BaseAgent(BaseModel, ABC, metaclass=AgentMeta):
         default=None,
         description="Knowledge sources for the agent.",
     )
-    knowledge_storage: InstanceOf[BaseKnowledgeStorage] | None = Field(
+    knowledge_storage: BaseKnowledgeStorage | None = Field(
         default=None,
         description="Custom knowledge storage for the agent.",
     )

lib/crewai/src/crewai/cli/plus_api.py

@@ -73,6 +73,7 @@ class PlusAPI:
         description: str | None,
         encoded_file: str,
         available_exports: list[dict[str, Any]] | None = None,
+        tools_metadata: list[dict[str, Any]] | None = None,
     ) -> httpx.Response:
         params = {
             "handle": handle,
@@ -81,6 +82,9 @@ class PlusAPI:
             "file": encoded_file,
             "description": description,
             "available_exports": available_exports,
+            "tools_metadata": {"package": handle, "tools": tools_metadata}
+            if tools_metadata is not None
+            else None,
         }
         return self._make_request("POST", f"{self.TOOLS_RESOURCE}", json=params)
 

lib/crewai/src/crewai/cli/templates/crew/pyproject.toml

@@ -5,7 +5,7 @@ description = "{{name}} using crewAI"
 authors = [{ name = "Your Name", email = "you@example.com" }]
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "crewai[tools]==1.12.2"
+    "crewai[tools]==1.13.0a5"
 ]
 
 [project.scripts]

lib/crewai/src/crewai/cli/templates/flow/pyproject.toml

@@ -5,7 +5,7 @@ description = "{{name}} using crewAI"
 authors = [{ name = "Your Name", email = "you@example.com" }]
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "crewai[tools]==1.12.2"
+    "crewai[tools]==1.13.0a5"
 ]
 
 [project.scripts]

lib/crewai/src/crewai/cli/templates/tool/pyproject.toml

@@ -5,7 +5,7 @@ description = "Power up your crews with {{folder_name}}"
 readme = "README.md"
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "crewai[tools]==1.12.2"
+    "crewai[tools]==1.13.0a5"
 ]
 
 [tool.crewai]

lib/crewai/src/crewai/cli/tools/main.py

@@ -17,6 +17,7 @@ from crewai.cli.constants import DEFAULT_CREWAI_ENTERPRISE_URL
 from crewai.cli.utils import (
     build_env_with_tool_repository_credentials,
     extract_available_exports,
+    extract_tools_metadata,
     get_project_description,
     get_project_name,
     get_project_version,
@@ -101,6 +102,18 @@ class ToolCommand(BaseCommand, PlusAPIMixin):
             console.print(
                 f"[green]Found these tools to publish: {', '.join([e['name'] for e in available_exports])}[/green]"
             )
+
+        console.print("[bold blue]Extracting tool metadata...[/bold blue]")
+        try:
+            tools_metadata = extract_tools_metadata()
+        except Exception as e:
+            console.print(
+                f"[yellow]Warning: Could not extract tool metadata: {e}[/yellow]\n"
+                f"Publishing will continue without detailed metadata."
+            )
+            tools_metadata = []
+
+        self._print_tools_preview(tools_metadata)
         self._print_current_organization()
 
         with tempfile.TemporaryDirectory() as temp_build_dir:
@@ -118,7 +131,7 @@ class ToolCommand(BaseCommand, PlusAPIMixin):
                     "Project build failed. Please ensure that the command `uv build --sdist` completes successfully.",
                     style="bold red",
                 )
-                raise SystemExit
+                raise SystemExit(1)
 
             tarball_path = os.path.join(temp_build_dir, tarball_filename)
             with open(tarball_path, "rb") as file:
@@ -134,6 +147,7 @@ class ToolCommand(BaseCommand, PlusAPIMixin):
             description=project_description,
             encoded_file=f"data:application/x-gzip;base64,{encoded_tarball}",
             available_exports=available_exports,
+            tools_metadata=tools_metadata,
         )
 
         self._validate_response(publish_response)
@@ -246,6 +260,55 @@ class ToolCommand(BaseCommand, PlusAPIMixin):
             )
             raise SystemExit
 
+    def _print_tools_preview(self, tools_metadata: list[dict[str, Any]]) -> None:
+        if not tools_metadata:
+            console.print("[yellow]No tool metadata extracted.[/yellow]")
+            return
+
+        console.print(
+            f"\n[bold]Tools to be published ({len(tools_metadata)}):[/bold]\n"
+        )
+
+        for tool in tools_metadata:
+            console.print(f"  [bold cyan]{tool.get('name', 'Unknown')}[/bold cyan]")
+            if tool.get("module"):
+                console.print(f"    Module: {tool.get('module')}")
+            console.print(f"    Name: {tool.get('humanized_name', 'N/A')}")
+            console.print(
+                f"    Description: {tool.get('description', 'N/A')[:80]}{'...' if len(tool.get('description', '')) > 80 else ''}"
+            )
+
+            init_params = tool.get("init_params_schema", {}).get("properties", {})
+            if init_params:
+                required = tool.get("init_params_schema", {}).get("required", [])
+                console.print("    Init parameters:")
+                for param_name, param_info in init_params.items():
+                    param_type = param_info.get("type", "any")
+                    is_required = param_name in required
+                    req_marker = "[red]*[/red]" if is_required else ""
+                    default = (
+                        f" = {param_info['default']}" if "default" in param_info else ""
+                    )
+                    console.print(
+                        f"      - {param_name}: {param_type}{default} {req_marker}"
+                    )
+
+            env_vars = tool.get("env_vars", [])
+            if env_vars:
+                console.print("    Environment variables:")
+                for env_var in env_vars:
+                    req_marker = "[red]*[/red]" if env_var.get("required") else ""
+                    default = (
+                        f" (default: {env_var['default']})"
+                        if env_var.get("default")
+                        else ""
+                    )
+                    console.print(
+                        f"      - {env_var['name']}: {env_var.get('description', 'N/A')}{default} {req_marker}"
+                    )
+
+            console.print()
+
     def _print_current_organization(self) -> None:
         settings = Settings()
         if settings.org_uuid:

lib/crewai/src/crewai/cli/utils.py

@@ -1,10 +1,15 @@
-from functools import reduce
+from collections.abc import Generator, Mapping
+from contextlib import contextmanager
+from functools import lru_cache, reduce
+import hashlib
 import importlib.util
+import inspect
 from inspect import getmro, isclass, isfunction, ismethod
 import os
 from pathlib import Path
 import shutil
 import sys
+import types
 from typing import Any, cast, get_type_hints
 
 import click
@@ -544,43 +549,62 @@ def build_env_with_tool_repository_credentials(
     return env
 
 
+@contextmanager
+def _load_module_from_file(
+    init_file: Path, module_name: str | None = None
+) -> Generator[types.ModuleType | None, None, None]:
+    """
+    Context manager for loading a module from file with automatic cleanup.
+
+    Yields the loaded module or None if loading fails.
+    """
+    if module_name is None:
+        module_name = (
+            f"temp_module_{hashlib.sha256(str(init_file).encode()).hexdigest()[:8]}"
+        )
+
+    spec = importlib.util.spec_from_file_location(module_name, init_file)
+    if not spec or not spec.loader:
+        yield None
+        return
+
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[module_name] = module
+
+    try:
+        spec.loader.exec_module(module)
+        yield module
+    finally:
+        sys.modules.pop(module_name, None)
+
+
 def _load_tools_from_init(init_file: Path) -> list[dict[str, Any]]:
     """
     Load and validate tools from a given __init__.py file.
     """
-    spec = importlib.util.spec_from_file_location("temp_module", init_file)
-
-    if not spec or not spec.loader:
-        return []
-
-    module = importlib.util.module_from_spec(spec)
-    sys.modules["temp_module"] = module
-
     try:
-        spec.loader.exec_module(module)
+        with _load_module_from_file(init_file) as module:
+            if module is None:
+                return []
 
-        if not hasattr(module, "__all__"):
-            console.print(
-                f"Warning: No __all__ defined in {init_file}",
-                style="bold yellow",
-            )
-            raise SystemExit(1)
-
-        return [
-            {
-                "name": name,
-            }
-            for name in module.__all__
-            if hasattr(module, name) and is_valid_tool(getattr(module, name))
-        ]
+            if not hasattr(module, "__all__"):
+                console.print(
+                    f"Warning: No __all__ defined in {init_file}",
+                    style="bold yellow",
+                )
+                raise SystemExit(1)
 
+            return [
+                {"name": name}
+                for name in module.__all__
+                if hasattr(module, name) and is_valid_tool(getattr(module, name))
+            ]
+    except SystemExit:
+        raise
     except Exception as e:
         console.print(f"[red]Warning: Could not load {init_file}: {e!s}[/red]")
         raise SystemExit(1) from e
 
-    finally:
-        sys.modules.pop("temp_module", None)
-
 
 def _print_no_tools_warning() -> None:
     """
@@ -610,3 +634,242 @@ def _print_no_tools_warning() -> None:
         "    # ... implementation\n"
         "    return result\n"
     )
+
+
+def extract_tools_metadata(dir_path: str = "src") -> list[dict[str, Any]]:
+    """
+    Extract rich metadata from tool classes in the project.
+
+    Returns a list of tool metadata dictionaries containing:
+    - name: Class name
+    - humanized_name: From name field default
+    - description: From description field default
+    - run_params_schema: JSON Schema for _run() params (from args_schema)
+    - init_params_schema: JSON Schema for __init__ params (filtered)
+    - env_vars: List of environment variable dicts
+    """
+    tools_metadata: list[dict[str, Any]] = []
+
+    for init_file in Path(dir_path).glob("**/__init__.py"):
+        tools = _extract_tool_metadata_from_init(init_file)
+        tools_metadata.extend(tools)
+
+    return tools_metadata
+
+
+def _extract_tool_metadata_from_init(init_file: Path) -> list[dict[str, Any]]:
+    """
+    Load module from init file and extract metadata from valid tool classes.
+    """
+    from crewai.tools.base_tool import BaseTool
+
+    try:
+        with _load_module_from_file(init_file) as module:
+            if module is None:
+                return []
+
+            exported_names = getattr(module, "__all__", None)
+            if not exported_names:
+                return []
+
+            tools_metadata = []
+            for name in exported_names:
+                obj = getattr(module, name, None)
+                if obj is None or not (
+                    inspect.isclass(obj) and issubclass(obj, BaseTool)
+                ):
+                    continue
+                if tool_info := _extract_single_tool_metadata(obj):
+                    tools_metadata.append(tool_info)
+
+            return tools_metadata
+    except Exception as e:
+        console.print(
+            f"[yellow]Warning: Could not extract metadata from {init_file}: {e}[/yellow]"
+        )
+        return []
+
+
+def _extract_single_tool_metadata(tool_class: type) -> dict[str, Any] | None:
+    """
+    Extract metadata from a single tool class.
+    """
+    try:
+        core_schema = cast(Any, tool_class).__pydantic_core_schema__
+        if not core_schema:
+            return None
+
+        schema = _unwrap_schema(core_schema)
+        fields = schema.get("schema", {}).get("fields", {})
+
+        try:
+            file_path = inspect.getfile(tool_class)
+            relative_path = Path(file_path).relative_to(Path.cwd())
+            module_path = relative_path.with_suffix("")
+            if module_path.parts[0] == "src":
+                module_path = Path(*module_path.parts[1:])
+            if module_path.name == "__init__":
+                module_path = module_path.parent
+            module = ".".join(module_path.parts)
+        except (TypeError, ValueError):
+            module = tool_class.__module__
+
+        return {
+            "name": tool_class.__name__,
+            "module": module,
+            "humanized_name": _extract_field_default(
+                fields.get("name"), fallback=tool_class.__name__
+            ),
+            "description": str(
+                _extract_field_default(fields.get("description"))
+            ).strip(),
+            "run_params_schema": _extract_run_params_schema(fields.get("args_schema")),
+            "init_params_schema": _extract_init_params_schema(tool_class),
+            "env_vars": _extract_env_vars(fields.get("env_vars")),
+        }
+
+    except Exception:
+        return None
+
+
+def _unwrap_schema(schema: Mapping[str, Any] | dict[str, Any]) -> dict[str, Any]:
+    """
+    Unwrap nested schema structures to get to the actual schema definition.
+    """
+    result: dict[str, Any] = dict(schema)
+    while (
+        result.get("type")
+        in {"function-after", "function-before", "function-wrap", "default"}
+        and "schema" in result
+    ):
+        result = dict(result["schema"])
+    if result.get("type") == "definitions" and "schema" in result:
+        result = dict(result["schema"])
+    return result
+
+
+def _extract_field_default(
+    field: dict[str, Any] | None, fallback: str | list[Any] = ""
+) -> str | list[Any] | int:
+    """
+    Extract the default value from a field schema.
+    """
+    if not field:
+        return fallback
+
+    schema = field.get("schema", {})
+    default = schema.get("default")
+    return default if isinstance(default, (list, str, int)) else fallback
+
+
+@lru_cache(maxsize=1)
+def _get_schema_generator() -> type:
+    """Get a SchemaGenerator that omits non-serializable defaults."""
+    from pydantic.json_schema import GenerateJsonSchema
+    from pydantic_core import PydanticOmit
+
+    class SchemaGenerator(GenerateJsonSchema):
+        def handle_invalid_for_json_schema(
+            self, schema: Any, error_info: Any
+        ) -> dict[str, Any]:
+            raise PydanticOmit
+
+    return SchemaGenerator
+
+
+def _extract_run_params_schema(
+    args_schema_field: dict[str, Any] | None,
+) -> dict[str, Any]:
+    """
+    Extract JSON Schema for the tool's run parameters from args_schema field.
+    """
+    from pydantic import BaseModel
+
+    if not args_schema_field:
+        return {}
+
+    args_schema_class = args_schema_field.get("schema", {}).get("default")
+    if not (
+        inspect.isclass(args_schema_class) and issubclass(args_schema_class, BaseModel)
+    ):
+        return {}
+
+    try:
+        return args_schema_class.model_json_schema(
+            schema_generator=_get_schema_generator()
+        )
+    except Exception:
+        return {}
+
+
+_IGNORED_INIT_PARAMS = frozenset(
+    {
+        "name",
+        "description",
+        "env_vars",
+        "args_schema",
+        "description_updated",
+        "cache_function",
+        "result_as_answer",
+        "max_usage_count",
+        "current_usage_count",
+        "package_dependencies",
+    }
+)
+
+
+def _extract_init_params_schema(tool_class: type) -> dict[str, Any]:
+    """
+    Extract JSON Schema for the tool's __init__ parameters, filtering out base fields.
+    """
+    try:
+        json_schema: dict[str, Any] = cast(Any, tool_class).model_json_schema(
+            schema_generator=_get_schema_generator(), mode="serialization"
+        )
+        filtered_properties = {
+            key: value
+            for key, value in json_schema.get("properties", {}).items()
+            if key not in _IGNORED_INIT_PARAMS
+        }
+        json_schema["properties"] = filtered_properties
+        if "required" in json_schema:
+            json_schema["required"] = [
+                key for key in json_schema["required"] if key in filtered_properties
+            ]
+        return json_schema
+    except Exception:
+        return {}
+
+
+def _extract_env_vars(env_vars_field: dict[str, Any] | None) -> list[dict[str, Any]]:
+    """
+    Extract environment variable definitions from env_vars field.
+    """
+    from crewai.tools.base_tool import EnvVar
+
+    if not env_vars_field:
+        return []
+
+    schema = env_vars_field.get("schema", {})
+    default = schema.get("default")
+    if default is None:
+        default_factory = schema.get("default_factory")
+        if callable(default_factory):
+            try:
+                default = default_factory()
+            except Exception:
+                default = []
+
+    if not isinstance(default, list):
+        return []
+
+    return [
+        {
+            "name": env_var.name,
+            "description": env_var.description,
+            "required": env_var.required,
+            "default": env_var.default,
+        }
+        for env_var in default
+        if isinstance(env_var, EnvVar)
+    ]

lib/crewai/src/crewai/crew.py

@@ -22,7 +22,6 @@ from pydantic import (
     UUID4,
     BaseModel,
     Field,
-    InstanceOf,
     Json,
     PrivateAttr,
     field_validator,
@@ -176,7 +175,7 @@ class Crew(FlowTrackable, BaseModel):
     _rpm_controller: RPMController = PrivateAttr()
     _logger: Logger = PrivateAttr()
     _file_handler: FileHandler = PrivateAttr()
-    _cache_handler: InstanceOf[CacheHandler] = PrivateAttr(default_factory=CacheHandler)
+    _cache_handler: CacheHandler = PrivateAttr(default_factory=CacheHandler)
     _memory: Memory | MemoryScope | MemorySlice | None = PrivateAttr(default=None)
     _train: bool | None = PrivateAttr(default=False)
     _train_iteration: int | None = PrivateAttr()
@@ -210,13 +209,13 @@ class Crew(FlowTrackable, BaseModel):
         default=None,
         description="Metrics for the LLM usage during all tasks execution.",
     )
-    manager_llm: str | InstanceOf[BaseLLM] | None = Field(
+    manager_llm: str | BaseLLM | None = Field(
         description="Language model that will run the agent.", default=None
     )
     manager_agent: BaseAgent | None = Field(
         description="Custom agent that will be used as manager.", default=None
     )
-    function_calling_llm: str | InstanceOf[LLM] | None = Field(
+    function_calling_llm: str | LLM | None = Field(
         description="Language model that will run the agent.", default=None
     )
     config: Json[dict[str, Any]] | dict[str, Any] | None = Field(default=None)
@@ -267,7 +266,7 @@ class Crew(FlowTrackable, BaseModel):
         default=False,
         description="Plan the crew execution and add the plan to the crew.",
     )
-    planning_llm: str | InstanceOf[BaseLLM] | Any | None = Field(
+    planning_llm: str | BaseLLM | Any | None = Field(
         default=None,
         description=(
             "Language model that will run the AgentPlanner if planning is True."
@@ -288,7 +287,7 @@ class Crew(FlowTrackable, BaseModel):
             "knowledge object."
         ),
     )
-    chat_llm: str | InstanceOf[BaseLLM] | Any | None = Field(
+    chat_llm: str | BaseLLM | Any | None = Field(
         default=None,
         description="LLM used to handle chatting with the crew.",
     )
@@ -1800,7 +1799,7 @@ class Crew(FlowTrackable, BaseModel):
     def test(
         self,
         n_iterations: int,
-        eval_llm: str | InstanceOf[BaseLLM],
+        eval_llm: str | BaseLLM,
         inputs: dict[str, Any] | None = None,
     ) -> None:
         """Test and evaluate the Crew with the given inputs for n iterations.

lib/crewai/src/crewai/events/types/llm_events.py

@@ -57,6 +57,7 @@ class LLMCallCompletedEvent(LLMEventBase):
     messages: str | list[dict[str, Any]] | None = None
     response: Any
     call_type: LLMCallType
+    usage: dict[str, Any] | None = None
 
 
 class LLMCallFailedEvent(LLMEventBase):

lib/crewai/src/crewai/experimental/agent_executor.py

@@ -11,10 +11,15 @@ import threading
 from typing import TYPE_CHECKING, Any, Literal, TypeVar, cast
 from uuid import uuid4
 
-from pydantic import BaseModel, Field, GetCoreSchemaHandler
-from pydantic_core import CoreSchema, core_schema
+from pydantic import (
+    BaseModel,
+    Field,
+    PrivateAttr,
+    model_validator,
+)
 from rich.console import Console
 from rich.text import Text
+from typing_extensions import Self
 
 from crewai.agents.agent_builder.base_agent_executor_mixin import CrewAgentExecutorMixin
 from crewai.agents.parser import (
@@ -119,6 +124,7 @@ class AgentExecutorState(BaseModel):
     (todos, observations, replan tracking) in a single validated model.
     """
 
+    id: str = Field(default_factory=lambda: str(uuid4()))
     messages: list[LLMMessage] = Field(default_factory=list)
     iterations: int = Field(default=0)
     current_answer: AgentAction | AgentFinish | None = Field(default=None)
@@ -152,6 +158,9 @@ class AgentExecutorState(BaseModel):
 class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
     """Agent Executor for both standalone agents and crew-bound agents.
 
+    _skip_auto_memory prevents Flow from eagerly allocating a Memory
+    instance — the executor uses agent/crew memory, not its own.
+
     Inherits from:
     - Flow[AgentExecutorState]: Provides flow orchestration capabilities
     - CrewAgentExecutorMixin: Provides memory methods (short/long/external term)
@@ -159,136 +168,74 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
     This executor can operate in two modes:
     - Standalone mode: When crew and task are None (used by Agent.kickoff())
     - Crew mode: When crew and task are provided (used by Agent.execute_task())
-
-    Note: Multiple instances may be created during agent initialization
-    (cache setup, RPM controller setup, etc.) but only the final instance
-    should execute tasks via invoke().
     """
 
-    def __init__(
-        self,
-        llm: BaseLLM,
-        agent: Agent,
-        prompt: SystemPromptResult | StandardPromptResult,
-        max_iter: int,
-        tools: list[CrewStructuredTool],
-        tools_names: str,
-        stop_words: list[str],
-        tools_description: str,
-        tools_handler: ToolsHandler,
-        task: Task | None = None,
-        crew: Crew | None = None,
-        step_callback: Any = None,
-        original_tools: list[BaseTool] | None = None,
-        function_calling_llm: BaseLLM | Any | None = None,
-        respect_context_window: bool = False,
-        request_within_rpm_limit: Callable[[], bool] | None = None,
-        callbacks: list[Any] | None = None,
-        response_model: type[BaseModel] | None = None,
-        i18n: I18N | None = None,
-    ) -> None:
-        """Initialize the flow-based agent executor.
+    _skip_auto_memory: bool = True
 
-        Args:
-            llm: Language model instance.
-            agent: Agent to execute.
-            prompt: Prompt templates.
-            max_iter: Maximum iterations.
-            tools: Available tools.
-            tools_names: Tool names string.
-            stop_words: Stop word list.
-            tools_description: Tool descriptions.
-            tools_handler: Tool handler instance.
-            task: Optional task to execute (None for standalone agent execution).
-            crew: Optional crew instance (None for standalone agent execution).
-            step_callback: Optional step callback.
-            original_tools: Original tool list.
-            function_calling_llm: Optional function calling LLM.
-            respect_context_window: Respect context limits.
-            request_within_rpm_limit: RPM limit check function.
-            callbacks: Optional callbacks list.
-            response_model: Optional Pydantic model for structured outputs.
-        """
-        self._i18n: I18N = i18n or get_i18n()
-        self.llm = llm
-        self.task: Task | None = task
-        self.agent = agent
-        self.crew: Crew | None = crew
-        self.prompt = prompt
-        self.tools = tools
-        self.tools_names = tools_names
-        self.stop = stop_words
-        self.max_iter = max_iter
-        self.callbacks = callbacks or []
-        self._printer: Printer = Printer()
-        self.tools_handler = tools_handler
-        self.original_tools = original_tools or []
-        self.step_callback = step_callback
-        self.tools_description = tools_description
-        self.function_calling_llm = function_calling_llm
-        self.respect_context_window = respect_context_window
-        self.request_within_rpm_limit = request_within_rpm_limit
-        self.response_model = response_model
-        self.log_error_after = 3
-        self._console: Console = Console()
+    suppress_flow_events: bool = True  # always suppress for executor
+    llm: BaseLLM = Field(exclude=True)
+    agent: Agent = Field(exclude=True)
+    prompt: SystemPromptResult | StandardPromptResult = Field(exclude=True)
+    max_iter: int = Field(default=25, exclude=True)
+    tools: list[CrewStructuredTool] = Field(default_factory=list, exclude=True)
+    tools_names: str = Field(default="", exclude=True)
+    stop_words: list[str] = Field(default_factory=list, exclude=True)
+    tools_description: str = Field(default="", exclude=True)
+    tools_handler: ToolsHandler | None = Field(default=None, exclude=True)
+    task: Task | None = Field(default=None, exclude=True)
+    crew: Crew | None = Field(default=None, exclude=True)
+    step_callback: Any = Field(default=None, exclude=True)
+    original_tools: list[BaseTool] = Field(default_factory=list, exclude=True)
+    function_calling_llm: BaseLLM | None = Field(default=None, exclude=True)
+    respect_context_window: bool = Field(default=False, exclude=True)
+    request_within_rpm_limit: Callable[[], bool] | None = Field(
+        default=None, exclude=True
+    )
+    callbacks: list[Any] = Field(default_factory=list, exclude=True)
+    response_model: type[BaseModel] | None = Field(default=None, exclude=True)
+    i18n: I18N | None = Field(default=None, exclude=True)
+    log_error_after: int = Field(default=3, exclude=True)
+    before_llm_call_hooks: list[BeforeLLMCallHookType | BeforeLLMCallHookCallable] = (
+        Field(default_factory=list, exclude=True)
+    )
+    after_llm_call_hooks: list[AfterLLMCallHookType | AfterLLMCallHookCallable] = Field(
+        default_factory=list, exclude=True
+    )
 
-        # Error context storage for recovery
-        self._last_parser_error: OutputParserError | None = None
-        self._last_context_error: Exception | None = None
+    _i18n: I18N = PrivateAttr(default_factory=get_i18n)
+    _printer: Printer = PrivateAttr(default_factory=Printer)
+    _console: Console = PrivateAttr(default_factory=Console)
+    _last_parser_error: OutputParserError | None = PrivateAttr(default=None)
+    _last_context_error: Exception | None = PrivateAttr(default=None)
+    _execution_lock: threading.Lock = PrivateAttr(default_factory=threading.Lock)
+    _finalize_lock: threading.Lock = PrivateAttr(default_factory=threading.Lock)
+    _finalize_called: bool = PrivateAttr(default=False)
+    _is_executing: bool = PrivateAttr(default=False)
+    _has_been_invoked: bool = PrivateAttr(default=False)
+    _instance_id: str = PrivateAttr(default_factory=lambda: str(uuid4())[:8])
+    _step_executor: Any = PrivateAttr(default=None)
+    _planner_observer: Any = PrivateAttr(default=None)
 
-        # Execution guard to prevent concurrent/duplicate executions
-        self._execution_lock = threading.Lock()
-        self._finalize_lock = threading.Lock()
-        self._finalize_called: bool = False
-        self._is_executing: bool = False
-        self._has_been_invoked: bool = False
-        self._flow_initialized: bool = False
-
-        self._instance_id = str(uuid4())[:8]
-
-        self.before_llm_call_hooks: list[
-            BeforeLLMCallHookType | BeforeLLMCallHookCallable
-        ] = []
-        self.after_llm_call_hooks: list[
-            AfterLLMCallHookType | AfterLLMCallHookCallable
-        ] = []
+    @model_validator(mode="after")
+    def _setup_executor(self) -> Self:
+        """Configure executor after Pydantic field initialization."""
+        self._i18n = self.i18n or get_i18n()
         self.before_llm_call_hooks.extend(get_before_llm_call_hooks())
         self.after_llm_call_hooks.extend(get_after_llm_call_hooks())
 
         if self.llm:
             existing_stop = getattr(self.llm, "stop", [])
-            self.llm.stop = list(
-                set(
-                    existing_stop + self.stop
-                    if isinstance(existing_stop, list)
-                    else self.stop
-                )
-            )
+            if not isinstance(existing_stop, list):
+                existing_stop = []
+            self.llm.stop = list(set(existing_stop + self.stop_words))
+
         self._state = AgentExecutorState()
+        self.max_method_calls = self.max_iter * 10
 
-        # Plan-and-Execute components (Phase 2)
-        # Lazy-imported to avoid circular imports during module load
-        self._step_executor: Any = None
-        self._planner_observer: Any = None
-
-    def _ensure_flow_initialized(self) -> None:
-        """Ensure Flow.__init__() has been called.
-
-        This is deferred from __init__ to prevent FlowCreatedEvent emission
-        during agent setup when multiple executor instances are created.
-        Only the instance that actually executes via invoke() will emit events.
-        """
-        if not self._flow_initialized:
-            current_tracing = is_tracing_enabled_in_context()
-            # Now call Flow's __init__ which will replace self._state
-            # with Flow's managed state. Suppress flow events since this is
-            # an agent executor, not a user-facing flow.
-            super().__init__(
-                suppress_flow_events=True,
-                tracing=current_tracing if current_tracing else None,
-                max_method_calls=self.max_iter * 10,
-            )
-            self._flow_initialized = True
+        current_tracing = is_tracing_enabled_in_context()
+        self.tracing = current_tracing if current_tracing else None
+        self._flow_post_init()
+        return self
 
     def _check_native_tool_support(self) -> bool:
         """Check if LLM supports native function calling."""
@@ -318,19 +265,13 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
 
     @property
     def state(self) -> AgentExecutorState:
-        """Get state - returns temporary state if Flow not yet initialized.
-
-        Flow initialization is deferred to prevent event emission during agent setup.
-        Returns the temporary state until invoke() is called.
-        """
-        if self._flow_initialized and hasattr(self, "_state_lock"):
-            return StateProxy(self._state, self._state_lock)  # type: ignore[return-value]
-        return self._state
+        """Get thread-safe state proxy."""
+        return StateProxy(self._state, self._state_lock)  # type: ignore[return-value]
 
     @property
     def iterations(self) -> int:
         """Compatibility property for mixin - returns state iterations."""
-        return self._state.iterations
+        return self._state.iterations  # type: ignore[no-any-return]
 
     @iterations.setter
     def iterations(self, value: int) -> None:
@@ -340,7 +281,7 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
     @property
     def messages(self) -> list[LLMMessage]:
         """Compatibility property - returns state messages."""
-        return self._state.messages
+        return self._state.messages  # type: ignore[no-any-return]
 
     @messages.setter
     def messages(self, value: list[LLMMessage]) -> None:
@@ -1966,42 +1907,10 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
             "original_tool": original_tool,
         }
 
-    def _extract_tool_name(self, tool_call: Any) -> str:
-        """Extract tool name from various tool call formats."""
-        if hasattr(tool_call, "function"):
-            return sanitize_tool_name(tool_call.function.name)
-        if hasattr(tool_call, "function_call") and tool_call.function_call:
-            return sanitize_tool_name(tool_call.function_call.name)
-        if hasattr(tool_call, "name"):
-            return sanitize_tool_name(tool_call.name)
-        if isinstance(tool_call, dict):
-            func_info = tool_call.get("function", {})
-            return sanitize_tool_name(
-                func_info.get("name", "") or tool_call.get("name", "unknown")
-            )
-        return "unknown"
-
-    @router(execute_native_tool)
-    def check_native_todo_completion(
-        self,
-    ) -> Literal["todo_satisfied", "todo_not_satisfied"]:
-        """Check if the native tool execution satisfied the active todo.
-
-        Similar to check_todo_completion but for native tool execution path.
-        """
-        current_todo = self.state.todos.current_todo
-
-        if not current_todo:
-            return "todo_not_satisfied"
-
-        # For native tools, any tool execution satisfies the todo
-        return "todo_satisfied"
-
     @listen("initialized")
     def continue_iteration(self) -> Literal["check_iteration"]:
         """Bridge listener that connects iteration loop back to iteration check."""
-        if self._flow_initialized:
-            self._discard_or_listener(FlowMethodName("continue_iteration"))
+        self._discard_or_listener(FlowMethodName("continue_iteration"))
         return "check_iteration"
 
     @router(or_(initialize_reasoning, continue_iteration))
@@ -2629,8 +2538,6 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
         if is_inside_event_loop():
             return self.invoke_async(inputs)
 
-        self._ensure_flow_initialized()
-
         with self._execution_lock:
             if self._is_executing:
                 raise RuntimeError(
@@ -2721,8 +2628,6 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
         Returns:
             Dictionary with agent output.
         """
-        self._ensure_flow_initialized()
-
         with self._execution_lock:
             if self._is_executing:
                 raise RuntimeError(
@@ -3038,17 +2943,6 @@ class AgentExecutor(Flow[AgentExecutorState], CrewAgentExecutorMixin):
         """
         return bool(self.crew and self.crew._train)
 
-    @classmethod
-    def __get_pydantic_core_schema__(
-        cls, _source_type: Any, _handler: GetCoreSchemaHandler
-    ) -> CoreSchema:
-        """Generate Pydantic core schema for Protocol compatibility.
-
-        Allows the executor to be used in Pydantic models without
-        requiring arbitrary_types_allowed=True.
-        """
-        return core_schema.any_schema()
-
 
 # Backward compatibility alias (deprecated)
 CrewAgentExecutorFlow = AgentExecutor

lib/crewai/src/crewai/flow/flow.py

@@ -39,7 +39,14 @@ from uuid import uuid4
 
 from opentelemetry import baggage
 from opentelemetry.context import attach, detach
-from pydantic import BaseModel, Field, ValidationError
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    PrivateAttr,
+    ValidationError,
+)
+from pydantic._internal._model_construction import ModelMetaclass
 from rich.console import Console
 from rich.panel import Panel
 
@@ -81,6 +88,7 @@ from crewai.flow.flow_wrappers import (
     SimpleFlowCondition,
     StartMethod,
 )
+from crewai.flow.human_feedback import HumanFeedbackResult
 from crewai.flow.input_provider import InputProvider
 from crewai.flow.persistence.base import FlowPersistence
 from crewai.flow.types import (
@@ -108,7 +116,6 @@ if TYPE_CHECKING:
     from crewai_files import FileInput
 
     from crewai.flow.async_feedback.types import PendingFeedbackContext
-    from crewai.flow.human_feedback import HumanFeedbackResult
     from crewai.llms.base_llm import BaseLLM
 
 from crewai.flow.visualization import build_flow_structure, render_interactive
@@ -728,7 +735,7 @@ class StateProxy(Generic[T]):
         return result
 
 
-class FlowMeta(type):
+class FlowMeta(ModelMetaclass):
     def __new__(
         mcs,
         name: str,
@@ -736,6 +743,45 @@ class FlowMeta(type):
         namespace: dict[str, Any],
         **kwargs: Any,
     ) -> type:
+        parent_fields: set[str] = set()
+        for base in bases:
+            if hasattr(base, "model_fields"):
+                parent_fields.update(base.model_fields)
+
+        annotations = namespace.get("__annotations__", {})
+        _skip_types = (classmethod, staticmethod, property)
+
+        for base in bases:
+            if isinstance(base, ModelMetaclass):
+                continue
+            for attr_name in getattr(base, "__annotations__", {}):
+                if attr_name not in annotations and attr_name not in namespace:
+                    annotations[attr_name] = ClassVar
+
+        for attr_name, attr_value in namespace.items():
+            if isinstance(attr_value, property) and attr_name not in annotations:
+                for base in bases:
+                    base_ann = getattr(base, "__annotations__", {})
+                    if attr_name in base_ann:
+                        annotations[attr_name] = ClassVar
+
+        for attr_name, attr_value in list(namespace.items()):
+            if attr_name in annotations or attr_name.startswith("_"):
+                continue
+            if attr_name in parent_fields:
+                annotations[attr_name] = Any
+                if isinstance(attr_value, BaseModel):
+                    namespace[attr_name] = Field(
+                        default_factory=lambda v=attr_value: v, exclude=True
+                    )
+                continue
+            if callable(attr_value) or isinstance(
+                attr_value, (*_skip_types, FlowMethod)
+            ):
+                continue
+            annotations[attr_name] = ClassVar[type(attr_value)]
+        namespace["__annotations__"] = annotations
+
         cls = super().__new__(mcs, name, bases, namespace)
 
         start_methods = []
@@ -820,88 +866,90 @@ class FlowMeta(type):
         return cls
 
 
-class Flow(Generic[T], metaclass=FlowMeta):
+class Flow(BaseModel, Generic[T], metaclass=FlowMeta):
     """Base class for all flows.
 
     type parameter T must be either dict[str, Any] or a subclass of BaseModel."""
 
+    model_config = ConfigDict(
+        arbitrary_types_allowed=True,
+        ignored_types=(StartMethod, ListenMethod, RouterMethod),
+        revalidate_instances="never",
+    )
+    __hash__ = object.__hash__
+
     _start_methods: ClassVar[list[FlowMethodName]] = []
     _listeners: ClassVar[dict[FlowMethodName, SimpleFlowCondition | FlowCondition]] = {}
     _routers: ClassVar[set[FlowMethodName]] = set()
     _router_paths: ClassVar[dict[FlowMethodName, list[FlowMethodName]]] = {}
-    initial_state: type[T] | T | None = None
-    name: str | None = None
-    tracing: bool | None = None
-    stream: bool = False
-    memory: Memory | MemoryScope | MemorySlice | None = None
-    input_provider: InputProvider | None = None
 
-    def __class_getitem__(cls: type[Flow[T]], item: type[T]) -> type[Flow[T]]:
-        class _FlowGeneric(cls):  # type: ignore
-            _initial_state_t = item
+    initial_state: Any = Field(default=None)
+    name: str | None = Field(default=None)
+    tracing: bool | None = Field(default=None)
+    stream: bool = Field(default=False)
+    memory: Memory | MemoryScope | MemorySlice | None = Field(default=None)
+    input_provider: InputProvider | None = Field(default=None)
+    suppress_flow_events: bool = Field(default=False)
+    human_feedback_history: list[HumanFeedbackResult] = Field(default_factory=list)
+    last_human_feedback: HumanFeedbackResult | None = Field(default=None)
+
+    persistence: Any = Field(default=None, exclude=True)
+    max_method_calls: int = Field(default=100, exclude=True)
+
+    _methods: dict[FlowMethodName, FlowMethod[Any, Any]] = PrivateAttr(
+        default_factory=dict
+    )
+    _method_execution_counts: dict[FlowMethodName, int] = PrivateAttr(
+        default_factory=dict
+    )
+    _pending_and_listeners: dict[PendingListenerKey, set[FlowMethodName]] = PrivateAttr(
+        default_factory=dict
+    )
+    _fired_or_listeners: set[FlowMethodName] = PrivateAttr(default_factory=set)
+    _method_outputs: list[Any] = PrivateAttr(default_factory=list)
+    _state_lock: threading.Lock = PrivateAttr(default_factory=threading.Lock)
+    _or_listeners_lock: threading.Lock = PrivateAttr(default_factory=threading.Lock)
+    _completed_methods: set[FlowMethodName] = PrivateAttr(default_factory=set)
+    _method_call_counts: dict[FlowMethodName, int] = PrivateAttr(default_factory=dict)
+    _is_execution_resuming: bool = PrivateAttr(default=False)
+    _event_futures: list[Future[None]] = PrivateAttr(default_factory=list)
+    _pending_feedback_context: PendingFeedbackContext | None = PrivateAttr(default=None)
+    _human_feedback_method_outputs: dict[str, Any] = PrivateAttr(default_factory=dict)
+    _input_history: list[InputHistoryEntry] = PrivateAttr(default_factory=list)
+    _state: Any = PrivateAttr(default=None)
+
+    def __class_getitem__(cls: type[Flow[T]], item: type[T]) -> type[Flow[T]]:  # type: ignore[override]
+        class _FlowGeneric(cls):  # type: ignore[valid-type,misc]
+            pass
 
         _FlowGeneric.__name__ = f"{cls.__name__}[{item.__name__}]"
+        _FlowGeneric._initial_state_t = item
         return _FlowGeneric
 
-    def __init__(
-        self,
-        persistence: FlowPersistence | None = None,
-        tracing: bool | None = None,
-        suppress_flow_events: bool = False,
-        max_method_calls: int = 100,
-        **kwargs: Any,
-    ) -> None:
-        """Initialize a new Flow instance.
+    def __setattr__(self, name: str, value: Any) -> None:
+        """Allow arbitrary attribute assignment for backward compat with plain class."""
+        if name in self.model_fields or name in self.__private_attributes__:
+            super().__setattr__(name, value)
+        else:
+            object.__setattr__(self, name, value)
 
-        Args:
-            persistence: Optional persistence backend for storing flow states
-            tracing: Whether to enable tracing. True=always enable, False=always disable, None=check environment/user settings
-            suppress_flow_events: Whether to suppress flow event emissions (internal use)
-            max_method_calls: Maximum times a single method can be called per execution before raising RecursionError
-            **kwargs: Additional state values to initialize or override
-        """
-        # Initialize basic instance attributes
-        self._methods: dict[FlowMethodName, FlowMethod[Any, Any]] = {}
-        self._method_execution_counts: dict[FlowMethodName, int] = {}
-        self._pending_and_listeners: dict[PendingListenerKey, set[FlowMethodName]] = {}
-        self._fired_or_listeners: set[FlowMethodName] = (
-            set()
-        )  # Track OR listeners that already fired
-        self._method_outputs: list[Any] = []  # list to store all method outputs
-        self._state_lock = threading.Lock()
-        self._or_listeners_lock = threading.Lock()
-        self._completed_methods: set[FlowMethodName] = (
-            set()
-        )  # Track completed methods for reload
-        self._method_call_counts: dict[FlowMethodName, int] = {}
-        self._max_method_calls = max_method_calls
-        self._persistence: FlowPersistence | None = persistence
-        self._is_execution_resuming: bool = False
-        self._event_futures: list[Future[None]] = []
+    def model_post_init(self, __context: Any) -> None:
+        self._flow_post_init()
 
-        # Human feedback storage
-        self.human_feedback_history: list[HumanFeedbackResult] = []
-        self.last_human_feedback: HumanFeedbackResult | None = None
-        self._pending_feedback_context: PendingFeedbackContext | None = None
-        # Per-method stash for real @human_feedback output (keyed by method name)
-        # Used to decouple routing outcome from method return value when emit is set
-        self._human_feedback_method_outputs: dict[str, Any] = {}
-        self.suppress_flow_events: bool = suppress_flow_events
+    def _flow_post_init(self) -> None:
+        """Heavy initialization: state creation, events, memory, method registration."""
+        if getattr(self, "_flow_post_init_done", False):
+            return
+        object.__setattr__(self, "_flow_post_init_done", True)
 
-        # User input history (for self.ask())
-        self._input_history: list[InputHistoryEntry] = []
+        if self._state is None:
+            self._state = self._create_initial_state()
 
-        # Initialize state with initial values
-        self._state = self._create_initial_state()
-        self.tracing = tracing
         tracing_enabled = should_enable_tracing(override=self.tracing)
         set_tracing_enabled(tracing_enabled)
 
         trace_listener = TraceCollectionListener()
         trace_listener.setup_listeners(crewai_event_bus)
-        # Apply any additional kwargs
-        if kwargs:
-            self._initialize_state(kwargs)
 
         if not self.suppress_flow_events:
             crewai_event_bus.emit(
@@ -1385,8 +1433,8 @@ class Flow(Generic[T], metaclass=FlowMeta):
         self._pending_feedback_context = None
 
         # Clear pending feedback from persistence
-        if self._persistence:
-            self._persistence.clear_pending_feedback(context.flow_id)
+        if self.persistence:
+            self.persistence.clear_pending_feedback(context.flow_id)
 
         # Emit feedback received event
         crewai_event_bus.emit(
@@ -1427,17 +1475,17 @@ class Flow(Generic[T], metaclass=FlowMeta):
             if isinstance(e, HumanFeedbackPending):
                 self._pending_feedback_context = e.context
 
-                if self._persistence is None:
+                if self.persistence is None:
                     from crewai.flow.persistence import SQLiteFlowPersistence
 
-                    self._persistence = SQLiteFlowPersistence()
+                    self.persistence = SQLiteFlowPersistence()
 
                 state_data = (
                     self._state
                     if isinstance(self._state, dict)
                     else self._state.model_dump()
                 )
-                self._persistence.save_pending_feedback(
+                self.persistence.save_pending_feedback(
                     flow_uuid=e.context.flow_id,
                     context=e.context,
                     state_data=state_data,
@@ -1487,39 +1535,33 @@ class Flow(Generic[T], metaclass=FlowMeta):
         """
         init_state = self.initial_state
 
-        # Handle case where initial_state is None but we have a type parameter
         if init_state is None and hasattr(self, "_initial_state_t"):
             state_type = self._initial_state_t
             if isinstance(state_type, type):
                 if issubclass(state_type, FlowState):
-                    # Create instance - FlowState auto-generates id via default_factory
                     instance = state_type()
-                    # Ensure id is set - generate UUID if empty
                     if not getattr(instance, "id", None):
                         object.__setattr__(instance, "id", str(uuid4()))
                     return cast(T, instance)
                 if issubclass(state_type, BaseModel):
-                    # Create a new type with FlowState first for proper id default
+
                     class StateWithId(FlowState, state_type):  # type: ignore
                         pass
 
                     instance = StateWithId()
-                    # Ensure id is set - generate UUID if empty
                     if not getattr(instance, "id", None):
                         object.__setattr__(instance, "id", str(uuid4()))
                     return cast(T, instance)
                 if state_type is dict:
                     return cast(T, {"id": str(uuid4())})
 
-        # Handle case where no initial state is provided
         if init_state is None:
             return cast(T, {"id": str(uuid4())})
 
-        # Handle case where initial_state is a type (class)
         if isinstance(init_state, type):
             state_class = init_state
             if issubclass(state_class, FlowState):
-                return state_class()
+                return cast(T, state_class())
             if issubclass(state_class, BaseModel):
                 model_fields = getattr(state_class, "model_fields", None)
                 if not model_fields or "id" not in model_fields:
@@ -1527,7 +1569,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
                 model_instance = state_class()
                 if not getattr(model_instance, "id", None):
                     object.__setattr__(model_instance, "id", str(uuid4()))
-                return model_instance
+                return cast(T, model_instance)
             if init_state is dict:
                 return cast(T, {"id": str(uuid4())})
 
@@ -1538,32 +1580,21 @@ class Flow(Generic[T], metaclass=FlowMeta):
                 new_state["id"] = str(uuid4())
             return cast(T, new_state)
 
-        # Handle BaseModel instance case
         if isinstance(init_state, BaseModel):
-            model = cast(BaseModel, init_state)
-            if not hasattr(model, "id"):
-                raise ValueError("Flow state model must have an 'id' field")
-
-            # Create new instance with same values to avoid mutations
-            if hasattr(model, "model_dump"):
-                # Pydantic v2
+            model = init_state
+            if hasattr(model, "id"):
                 state_dict = model.model_dump()
-            elif hasattr(model, "dict"):
-                # Pydantic v1
-                state_dict = model.dict()
-            else:
-                # Fallback for other BaseModel implementations
-                state_dict = {
-                    k: v for k, v in model.__dict__.items() if not k.startswith("_")
-                }
+                if not state_dict.get("id"):
+                    state_dict["id"] = str(uuid4())
+                model_class = type(model)
+                return cast(T, model_class(**state_dict))
 
-            # Ensure id is set - generate UUID if empty
-            if not state_dict.get("id"):
-                state_dict["id"] = str(uuid4())
+            class StateWithId(FlowState, type(model)):  # type: ignore
+                pass
 
-            # Create new instance of the same class
-            model_class = type(model)
-            return cast(T, model_class(**state_dict))
+            state_dict = model.model_dump()
+            state_dict["id"] = str(uuid4())
+            return cast(T, StateWithId(**state_dict))
         raise TypeError(
             f"Initial state must be dict or BaseModel, got {type(self.initial_state)}"
         )
@@ -1576,17 +1607,17 @@ class Flow(Generic[T], metaclass=FlowMeta):
         """
         if isinstance(self._state, BaseModel):
             try:
-                return self._state.model_copy(deep=True)
+                return cast(T, self._state.model_copy(deep=True))
             except (TypeError, AttributeError):
                 try:
                     state_dict = self._state.model_dump()
                     model_class = type(self._state)
-                    return model_class(**state_dict)
+                    return cast(T, model_class(**state_dict))
                 except Exception:
-                    return self._state.model_copy(deep=False)
+                    return cast(T, self._state.model_copy(deep=False))
         else:
             try:
-                return copy.deepcopy(self._state)
+                return cast(T, copy.deepcopy(self._state))
             except (TypeError, AttributeError):
                 return cast(T, self._state.copy())
 
@@ -1662,7 +1693,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
         elif isinstance(self._state, BaseModel):
             # For BaseModel states, preserve existing fields unless overridden
             try:
-                model = cast(BaseModel, self._state)
+                model = self._state
                 # Get current state as dict
                 if hasattr(model, "model_dump"):
                     current_state = model.model_dump()
@@ -1713,7 +1744,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
             self._state.update(stored_state)
         elif isinstance(self._state, BaseModel):
             # For BaseModel states, create new instance with stored values
-            model = cast(BaseModel, self._state)
+            model = self._state
             if hasattr(model, "model_validate"):
                 # Pydantic v2
                 self._state = cast(T, type(model).model_validate(stored_state))
@@ -1938,7 +1969,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
 
         try:
             # Reset flow state for fresh execution unless restoring from persistence
-            is_restoring = inputs and "id" in inputs and self._persistence is not None
+            is_restoring = inputs and "id" in inputs and self.persistence is not None
             if not is_restoring:
                 # Clear completed methods and outputs for a fresh start
                 self._completed_methods.clear()
@@ -1964,9 +1995,9 @@ class Flow(Generic[T], metaclass=FlowMeta):
                         setattr(self._state, "id", inputs["id"])  # noqa: B010
 
                 # If persistence is enabled, attempt to restore the stored state using the provided id.
-                if "id" in inputs and self._persistence is not None:
+                if "id" in inputs and self.persistence is not None:
                     restore_uuid = inputs["id"]
-                    stored_state = self._persistence.load_state(restore_uuid)
+                    stored_state = self.persistence.load_state(restore_uuid)
                     if stored_state:
                         self._log_flow_event(
                             f"Loading flow state from memory for UUID: {restore_uuid}"
@@ -2036,17 +2067,17 @@ class Flow(Generic[T], metaclass=FlowMeta):
 
                 if isinstance(e, HumanFeedbackPending):
                     # Auto-save pending feedback (create default persistence if needed)
-                    if self._persistence is None:
+                    if self.persistence is None:
                         from crewai.flow.persistence import SQLiteFlowPersistence
 
-                        self._persistence = SQLiteFlowPersistence()
+                        self.persistence = SQLiteFlowPersistence()
 
                     state_data = (
                         self._state
                         if isinstance(self._state, dict)
                         else self._state.model_dump()
                     )
-                    self._persistence.save_pending_feedback(
+                    self.persistence.save_pending_feedback(
                         flow_uuid=e.context.flow_id,
                         context=e.context,
                         state_data=state_data,
@@ -2332,10 +2363,10 @@ class Flow(Generic[T], metaclass=FlowMeta):
             if isinstance(e, HumanFeedbackPending):
                 e.context.method_name = method_name
 
-                if self._persistence is None:
+                if self.persistence is None:
                     from crewai.flow.persistence import SQLiteFlowPersistence
 
-                    self._persistence = SQLiteFlowPersistence()
+                    self.persistence = SQLiteFlowPersistence()
 
                 # Emit paused event (not failed)
                 if not self.suppress_flow_events:
@@ -2696,9 +2727,9 @@ class Flow(Generic[T], metaclass=FlowMeta):
             - Catches and logs any exceptions during execution, preventing individual listener failures from breaking the entire flow
         """
         count = self._method_call_counts.get(listener_name, 0) + 1
-        if count > self._max_method_calls:
+        if count > self.max_method_calls:
             raise RecursionError(
-                f"Method '{listener_name}' has been called {self._max_method_calls} times in "
+                f"Method '{listener_name}' has been called {self.max_method_calls} times in "
                 f"this flow execution, which indicates an infinite loop. "
                 f"This commonly happens when a @listen label matches the "
                 f"method's own name."
@@ -2805,7 +2836,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
 
         This is best-effort: if persistence is not configured, this is a no-op.
         """
-        if self._persistence is None:
+        if self.persistence is None:
             return
         try:
             state_data = (
@@ -2813,7 +2844,7 @@ class Flow(Generic[T], metaclass=FlowMeta):
                 if isinstance(self._state, dict)
                 else self._state.model_dump()
             )
-            self._persistence.save_state(
+            self.persistence.save_state(
                 flow_uuid=self.flow_id,
                 method_name="_ask_checkpoint",
                 state_data=state_data,

lib/crewai/src/crewai/knowledge/storage/base_knowledge_storage.py

@@ -3,12 +3,15 @@ from __future__ import annotations
 from abc import ABC, abstractmethod
 from typing import TYPE_CHECKING, Any
 
+from pydantic import BaseModel, ConfigDict
+
 
 if TYPE_CHECKING:
     from crewai.rag.types import SearchResult
 
 
-class BaseKnowledgeStorage(ABC):
+class BaseKnowledgeStorage(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
     """Abstract base class for knowledge storage implementations."""
 
     @abstractmethod

lib/crewai/src/crewai/knowledge/storage/knowledge_storage.py

@@ -3,6 +3,9 @@ import traceback
 from typing import Any, cast
 import warnings
 
+from pydantic import Field, PrivateAttr, model_validator
+from typing_extensions import Self
+
 from crewai.knowledge.storage.base_knowledge_storage import BaseKnowledgeStorage
 from crewai.rag.chromadb.config import ChromaDBConfig
 from crewai.rag.chromadb.types import ChromaEmbeddingFunctionWrapper
@@ -22,31 +25,32 @@ class KnowledgeStorage(BaseKnowledgeStorage):
     search efficiency.
     """
 
-    def __init__(
-        self,
-        embedder: ProviderSpec
+    collection_name: str | None = None
+    embedder: (
+        ProviderSpec
         | BaseEmbeddingsProvider[Any]
         | type[BaseEmbeddingsProvider[Any]]
-        | None = None,
-        collection_name: str | None = None,
-    ) -> None:
-        self.collection_name = collection_name
-        self._client: BaseClient | None = None
+        | None
+    ) = Field(default=None, exclude=True)
+    _client: BaseClient | None = PrivateAttr(default=None)
 
+    @model_validator(mode="after")
+    def _init_client(self) -> Self:
         warnings.filterwarnings(
             "ignore",
             message=r".*'model_fields'.*is deprecated.*",
             module=r"^chromadb(\.|$)",
         )
 
-        if embedder:
-            embedding_function = build_embedder(embedder)  # type: ignore[arg-type]
+        if self.embedder:
+            embedding_function = build_embedder(self.embedder)  # type: ignore[arg-type]
             config = ChromaDBConfig(
                 embedding_function=cast(
                     ChromaEmbeddingFunctionWrapper, embedding_function
                 )
             )
             self._client = create_client(config)
+        return self
 
     def _get_client(self) -> BaseClient:
         """Get the appropriate client - instance-specific or global."""

lib/crewai/src/crewai/lite_agent.py

@@ -22,7 +22,6 @@ from pydantic import (
     UUID4,
     BaseModel,
     Field,
-    InstanceOf,
     PrivateAttr,
     field_validator,
     model_validator,
@@ -204,7 +203,7 @@ class LiteAgent(FlowTrackable, BaseModel):
     role: str = Field(description="Role of the agent")
     goal: str = Field(description="Goal of the agent")
     backstory: str = Field(description="Backstory of the agent")
-    llm: str | InstanceOf[BaseLLM] | Any | None = Field(
+    llm: str | BaseLLM | Any | None = Field(
         default=None, description="Language model that will run the agent"
     )
     tools: list[BaseTool] = Field(

lib/crewai/src/crewai/llm.py

@@ -20,8 +20,7 @@ from typing import (
 )
 
 from dotenv import load_dotenv
-import httpx
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, model_validator
 from typing_extensions import Self
 
 from crewai.events.event_bus import crewai_event_bus
@@ -37,7 +36,12 @@ from crewai.events.types.tool_usage_events import (
     ToolUsageFinishedEvent,
     ToolUsageStartedEvent,
 )
-from crewai.llms.base_llm import BaseLLM, get_current_call_id, llm_call_context
+from crewai.llms.base_llm import (
+    BaseLLM,
+    JsonResponseFormat,
+    get_current_call_id,
+    llm_call_context,
+)
 from crewai.llms.constants import (
     ANTHROPIC_MODELS,
     AZURE_MODELS,
@@ -63,8 +67,6 @@ except ImportError:
 
 if TYPE_CHECKING:
     from crewai.agent.core import Agent
-    from crewai.llms.hooks.base import BaseInterceptor
-    from crewai.llms.providers.anthropic.completion import AnthropicThinkingConfig
     from crewai.task import Task
     from crewai.tools.base_tool import BaseTool
     from crewai.utilities.types import LLMMessage
@@ -342,6 +344,27 @@ class AccumulatedToolArgs(BaseModel):
 
 class LLM(BaseLLM):
     completion_cost: float | None = None
+    timeout: float | int | None = None
+    top_p: float | None = None
+    n: int | None = None
+    max_completion_tokens: int | None = None
+    max_tokens: int | float | None = None
+    presence_penalty: float | None = None
+    frequency_penalty: float | None = None
+    logit_bias: dict[int, float] | None = None
+    response_format: JsonResponseFormat | type[BaseModel] | None = None
+    seed: int | None = None
+    logprobs: int | None = None
+    top_logprobs: int | None = None
+    api_base: str | None = None
+    api_version: str | None = None
+    callbacks: list[Any] | None = None
+    reasoning_effort: Literal["none", "low", "medium", "high"] | None = None
+    stream: bool = False
+    interceptor: Any = None
+    thinking: Any = None
+    context_window_size: int = 0
+    is_anthropic: bool = False
 
     def __new__(cls, model: str, is_litellm: bool = False, **kwargs: Any) -> LLM:
         """Factory method that routes to native SDK or falls back to LiteLLM.
@@ -436,10 +459,7 @@ class LLM(BaseLLM):
             logger.error(error_msg)
             raise ImportError(error_msg) from None
 
-        instance = object.__new__(cls)
-        super(LLM, instance).__init__(model=model, is_litellm=True, **kwargs)
-        instance.is_litellm = True
-        return instance
+        return object.__new__(cls)
 
     @classmethod
     def _matches_provider_pattern(cls, model: str, provider: str) -> bool:
@@ -624,89 +644,23 @@ class LLM(BaseLLM):
 
         return None
 
-    def __init__(
-        self,
-        model: str,
-        timeout: float | int | None = None,
-        temperature: float | None = None,
-        top_p: float | None = None,
-        n: int | None = None,
-        stop: str | list[str] | None = None,
-        max_completion_tokens: int | None = None,
-        max_tokens: int | float | None = None,
-        presence_penalty: float | None = None,
-        frequency_penalty: float | None = None,
-        logit_bias: dict[int, float] | None = None,
-        response_format: type[BaseModel] | None = None,
-        seed: int | None = None,
-        logprobs: int | None = None,
-        top_logprobs: int | None = None,
-        base_url: str | None = None,
-        api_base: str | None = None,
-        api_version: str | None = None,
-        api_key: str | None = None,
-        callbacks: list[Any] | None = None,
-        reasoning_effort: Literal["none", "low", "medium", "high"] | None = None,
-        stream: bool = False,
-        interceptor: BaseInterceptor[httpx.Request, httpx.Response] | None = None,
-        thinking: AnthropicThinkingConfig | dict[str, Any] | None = None,
-        prefer_upload: bool = False,
-        **kwargs: Any,
-    ) -> None:
-        """Initialize LLM instance.
+    @model_validator(mode="before")
+    @classmethod
+    def _validate_llm_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+        model = data.get("model", "")
+        data["is_anthropic"] = cls._is_anthropic_model(model)
+        return data
 
-        Note: This __init__ method is only called for fallback instances.
-        Native provider instances handle their own initialization in their respective classes.
-        """
-        super().__init__(
-            model=model,
-            temperature=temperature,
-            api_key=api_key,
-            base_url=base_url,
-            timeout=timeout,
-            **kwargs,
-        )
-        self.model = model
-        self.timeout = timeout
-        self.temperature = temperature
-        self.top_p = top_p
-        self.n = n
-        self.max_completion_tokens = max_completion_tokens
-        self.max_tokens = max_tokens
-        self.presence_penalty = presence_penalty
-        self.frequency_penalty = frequency_penalty
-        self.logit_bias = logit_bias
-        self.response_format = response_format
-        self.seed = seed
-        self.logprobs = logprobs
-        self.top_logprobs = top_logprobs
-        self.base_url = base_url
-        self.api_base = api_base
-        self.api_version = api_version
-        self.api_key = api_key
-        self.callbacks = callbacks
-        self.context_window_size = 0
-        self.reasoning_effort = reasoning_effort
-        self.prefer_upload = prefer_upload
-        self.additional_params = {
-            k: v for k, v in kwargs.items() if k not in ("is_litellm", "provider")
-        }
-        self.is_anthropic = self._is_anthropic_model(model)
-        self.stream = stream
-        self.interceptor = interceptor
-
-        litellm.drop_params = True
-
-        # Normalize self.stop to always be a list[str]
-        if stop is None:
-            self.stop: list[str] = []
-        elif isinstance(stop, str):
-            self.stop = [stop]
-        else:
-            self.stop = stop
-
-        self.set_callbacks(callbacks or [])
-        self.set_env_callbacks()
+    @model_validator(mode="after")
+    def _init_litellm(self) -> LLM:
+        self.is_litellm = True
+        if LITELLM_AVAILABLE:
+            litellm.drop_params = True
+            self.set_callbacks(self.callbacks or [])
+            self.set_env_callbacks()
+        return self
 
     @staticmethod
     def _is_anthropic_model(model: str) -> bool:
@@ -753,7 +707,7 @@ class LLM(BaseLLM):
             "temperature": self.temperature,
             "top_p": self.top_p,
             "n": self.n,
-            "stop": self.stop or None,
+            "stop": (self.stop or None) if self.supports_stop_words() else None,
             "max_tokens": self.max_tokens or self.max_completion_tokens,
             "presence_penalty": self.presence_penalty,
             "frequency_penalty": self.frequency_penalty,
@@ -1016,21 +970,25 @@ class LLM(BaseLLM):
                     )
                     result = instructor_instance.to_pydantic()
                     structured_response = result.model_dump_json()
+                    usage_dict = self._usage_to_dict(usage_info)
                     self._handle_emit_call_events(
                         response=structured_response,
                         call_type=LLMCallType.LLM_CALL,
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage_dict,
                     )
                     return structured_response
 
+                usage_dict = self._usage_to_dict(usage_info)
                 self._handle_emit_call_events(
                     response=full_response,
                     call_type=LLMCallType.LLM_CALL,
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage_dict,
                 )
                 return full_response
 
@@ -1040,12 +998,14 @@ class LLM(BaseLLM):
                 return tool_result
 
             # --- 10) Emit completion event and return response
+            usage_dict = self._usage_to_dict(usage_info)
             self._handle_emit_call_events(
                 response=full_response,
                 call_type=LLMCallType.LLM_CALL,
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage_dict,
             )
             return full_response
 
@@ -1067,6 +1027,7 @@ class LLM(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=self._usage_to_dict(usage_info),
                 )
                 return full_response
 
@@ -1218,6 +1179,7 @@ class LLM(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=None,
             )
             return structured_response
 
@@ -1248,6 +1210,8 @@ class LLM(BaseLLM):
                 raise LLMContextLengthExceededError(error_msg) from e
             raise
 
+        response_usage = self._usage_to_dict(getattr(response, "usage", None))
+
         # --- 2) Handle structured output response (when response_model is provided)
         if response_model is not None:
             # When using instructor/response_model, litellm returns a Pydantic model instance
@@ -1259,6 +1223,7 @@ class LLM(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=response_usage,
                 )
                 return structured_response
 
@@ -1290,6 +1255,7 @@ class LLM(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=response_usage,
             )
             return text_response
 
@@ -1313,6 +1279,7 @@ class LLM(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=response_usage,
         )
         return text_response
 
@@ -1362,6 +1329,7 @@ class LLM(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=None,
             )
             return structured_response
 
@@ -1388,6 +1356,8 @@ class LLM(BaseLLM):
                 raise LLMContextLengthExceededError(error_msg) from e
             raise
 
+        response_usage = self._usage_to_dict(getattr(response, "usage", None))
+
         if response_model is not None:
             if isinstance(response, BaseModel):
                 structured_response = response.model_dump_json()
@@ -1397,6 +1367,7 @@ class LLM(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=response_usage,
                 )
                 return structured_response
 
@@ -1426,6 +1397,7 @@ class LLM(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=response_usage,
             )
             return text_response
 
@@ -1448,6 +1420,7 @@ class LLM(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=response_usage,
         )
         return text_response
 
@@ -1594,12 +1567,14 @@ class LLM(BaseLLM):
                     if result is not None:
                         return result
 
+            usage_dict = self._usage_to_dict(usage_info)
             self._handle_emit_call_events(
                 response=full_response,
                 call_type=LLMCallType.LLM_CALL,
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params.get("messages"),
+                usage=usage_dict,
             )
             return full_response
 
@@ -1621,6 +1596,7 @@ class LLM(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("messages"),
+                    usage=self._usage_to_dict(usage_info),
                 )
                 return full_response
             raise
@@ -1825,9 +1801,11 @@ class LLM(BaseLLM):
                     # whether to summarize the content or abort based on the respect_context_window flag
                     raise
                 except Exception as e:
-                    unsupported_stop = "Unsupported parameter" in str(
-                        e
-                    ) and "'stop'" in str(e)
+                    error_str = str(e)
+                    unsupported_stop = "'stop'" in error_str and (
+                        "Unsupported parameter" in error_str
+                        or "does not support parameters" in error_str
+                    )
 
                     if unsupported_stop:
                         if (
@@ -1961,9 +1939,11 @@ class LLM(BaseLLM):
                 except LLMContextLengthExceededError:
                     raise
                 except Exception as e:
-                    unsupported_stop = "Unsupported parameter" in str(
-                        e
-                    ) and "'stop'" in str(e)
+                    error_str = str(e)
+                    unsupported_stop = "'stop'" in error_str and (
+                        "Unsupported parameter" in error_str
+                        or "does not support parameters" in error_str
+                    )
 
                     if unsupported_stop:
                         if (
@@ -2003,6 +1983,19 @@ class LLM(BaseLLM):
                     )
                     raise
 
+    @staticmethod
+    def _usage_to_dict(usage: Any) -> dict[str, Any] | None:
+        if usage is None:
+            return None
+        if isinstance(usage, dict):
+            return usage
+        if hasattr(usage, "model_dump"):
+            result: dict[str, Any] = usage.model_dump()
+            return result
+        if hasattr(usage, "__dict__"):
+            return {k: v for k, v in vars(usage).items() if not k.startswith("_")}
+        return None
+
     def _handle_emit_call_events(
         self,
         response: Any,
@@ -2010,6 +2003,7 @@ class LLM(BaseLLM):
         from_task: Task | None = None,
         from_agent: Agent | None = None,
         messages: str | list[LLMMessage] | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> None:
         """Handle the events for the LLM call.
 
@@ -2019,6 +2013,7 @@ class LLM(BaseLLM):
             from_task: Optional task object
             from_agent: Optional agent object
             messages: Optional messages object
+            usage: Optional token usage data
         """
         crewai_event_bus.emit(
             self,
@@ -2030,6 +2025,7 @@ class LLM(BaseLLM):
                 from_agent=from_agent,
                 model=self.model,
                 call_id=get_current_call_id(),
+                usage=usage,
             ),
         )
 
@@ -2263,6 +2259,10 @@ class LLM(BaseLLM):
         Note: This method is only used by the litellm fallback path.
         Native providers override this method with their own implementation.
         """
+        model_lower = self.model.lower() if self.model else ""
+        if "gpt-5" in model_lower:
+            return False
+
         if not LITELLM_AVAILABLE or get_supported_openai_params is None:
             # When litellm is not available, assume stop words are supported
             return True
@@ -2434,7 +2434,7 @@ class LLM(BaseLLM):
             **filtered_params,
         )
 
-    def __deepcopy__(self, memo: dict[int, Any] | None) -> LLM:
+    def __deepcopy__(self, memo: dict[int, Any] | None = None) -> LLM:
         """Create a deep copy of the LLM instance."""
         import copy
 

lib/crewai/src/crewai/llms/base_llm.py

@@ -14,10 +14,18 @@ from datetime import datetime
 import json
 import logging
 import re
-from typing import TYPE_CHECKING, Any, Final
+from typing import TYPE_CHECKING, Any, Final, Literal
 import uuid
 
-from pydantic import BaseModel
+from pydantic import (
+    AliasChoices,
+    BaseModel,
+    ConfigDict,
+    Field,
+    PrivateAttr,
+    model_validator,
+)
+from typing_extensions import TypedDict
 
 from crewai.events.event_bus import crewai_event_bus
 from crewai.events.types.llm_events import (
@@ -51,6 +59,12 @@ if TYPE_CHECKING:
     from crewai.utilities.types import LLMMessage
 
 
+class JsonResponseFormat(TypedDict):
+    """Response format requesting raw JSON output (e.g. ``{"type": "json_object"}``)."""
+
+    type: Literal["json_object"]
+
+
 DEFAULT_CONTEXT_WINDOW_SIZE: Final[int] = 4096
 DEFAULT_SUPPORTS_STOP_WORDS: Final[bool] = True
 _JSON_EXTRACTION_PATTERN: Final[re.Pattern[str]] = re.compile(r"\{.*}", re.DOTALL)
@@ -82,7 +96,7 @@ def get_current_call_id() -> str:
     return call_id
 
 
-class BaseLLM(ABC):
+class BaseLLM(BaseModel, ABC):
     """Abstract base class for LLM implementations.
 
     This class defines the interface that all LLM implementations must follow.
@@ -101,56 +115,100 @@ class BaseLLM(ABC):
         additional_params: Additional provider-specific parameters.
     """
 
+    model_config = ConfigDict(arbitrary_types_allowed=True, populate_by_name=True)
+
+    model: str
+    temperature: float | None = None
+    api_key: str | None = None
+    base_url: str | None = None
+    provider: str = Field(default="openai")
+    prefer_upload: bool = False
     is_litellm: bool = False
+    stop: list[str] = Field(
+        default_factory=list,
+        validation_alias=AliasChoices("stop", "stop_sequences"),
+    )
+    additional_params: dict[str, Any] = Field(default_factory=dict)
 
-    def __init__(
-        self,
-        model: str,
-        temperature: float | None = None,
-        api_key: str | None = None,
-        base_url: str | None = None,
-        provider: str | None = None,
-        prefer_upload: bool = False,
-        **kwargs: Any,
-    ) -> None:
-        """Initialize the BaseLLM with default attributes.
+    def __setattr__(self, name: str, value: Any) -> None:
+        if name in ("stop", "stop_sequences"):
+            if value is None:
+                value = []
+            elif isinstance(value, str):
+                value = [value]
+            elif not isinstance(value, list):
+                value = list(value)
+            name = "stop"
+        try:
+            super().__setattr__(name, value)
+        except ValueError:
+            if name in self.model_fields:
+                raise  # Re-raise validation errors on declared fields
+            # Fallback for attributes not declared as fields (e.g. mock patching)
+            object.__setattr__(self, name, value)
+        except AttributeError:
+            object.__setattr__(self, name, value)
 
-        Args:
-            model: The model identifier/name.
-            temperature: Optional temperature setting for response generation.
-            stop: Optional list of stop sequences for generation.
-            prefer_upload: Whether to prefer file upload over inline base64.
-            **kwargs: Additional provider-specific parameters.
+    def __delattr__(self, name: str) -> None:
+        try:
+            super().__delattr__(name)
+        except AttributeError:
+            object.__delattr__(self, name)
+
+    @property
+    def stop_sequences(self) -> list[str]:
+        """Alias for ``stop`` — kept for backward compatibility with provider APIs.
+
+        Writes are handled by ``__setattr__``, which normalizes and redirects
+        ``stop_sequences`` assignments to the ``stop`` field.
         """
-        if not model:
-            raise ValueError("Model name is required and cannot be empty")
+        return self.stop
 
-        self.model = model
-        self.temperature = temperature
-        self.api_key = api_key
-        self.base_url = base_url
-        self.prefer_upload = prefer_upload
-        # Store additional parameters for provider-specific use
-        self.additional_params = kwargs
-        self._provider = provider or "openai"
-
-        stop = kwargs.pop("stop", None)
-        if stop is None:
-            self.stop: list[str] = []
-        elif isinstance(stop, str):
-            self.stop = [stop]
-        elif isinstance(stop, list):
-            self.stop = stop
-        else:
-            self.stop = []
-
-        self._token_usage = {
+    _token_usage: dict[str, int] = PrivateAttr(
+        default_factory=lambda: {
             "total_tokens": 0,
             "prompt_tokens": 0,
             "completion_tokens": 0,
             "successful_requests": 0,
             "cached_prompt_tokens": 0,
         }
+    )
+
+    @model_validator(mode="before")
+    @classmethod
+    def _validate_init_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+
+        if not data.get("model"):
+            raise ValueError("Model name is required and cannot be empty")
+
+        # Normalize stop: accept str, list, or None; also accept stop_sequences alias
+        stop_seqs = data.pop("stop_sequences", None)
+        stop = stop_seqs if stop_seqs is not None else data.get("stop")
+        if stop is None:
+            data["stop"] = []
+        elif isinstance(stop, str):
+            data["stop"] = [stop]
+        elif isinstance(stop, list):
+            data["stop"] = stop
+        else:
+            data["stop"] = list(stop)
+
+        # Default provider
+        if not data.get("provider"):
+            data["provider"] = "openai"
+
+        # Collect unknown kwargs into additional_params
+        known_fields = set(cls.model_fields.keys())
+        extras = {k: v for k, v in data.items() if k not in known_fields}
+        for k in extras:
+            data.pop(k)
+        existing = data.get("additional_params") or {}
+        existing.update(extras)
+        data["additional_params"] = existing
+
+        return data
 
     def to_config_dict(self) -> dict[str, Any]:
         """Serialize this LLM to a dict that can reconstruct it via ``LLM(**config)``.
@@ -174,16 +232,6 @@ class BaseLLM(ABC):
 
         return config
 
-    @property
-    def provider(self) -> str:
-        """Get the provider of the LLM."""
-        return self._provider
-
-    @provider.setter
-    def provider(self, value: str) -> None:
-        """Set the provider of the LLM."""
-        self._provider = value
-
     @abstractmethod
     def call(
         self,
@@ -412,6 +460,7 @@ class BaseLLM(ABC):
         from_task: Task | None = None,
         from_agent: Agent | None = None,
         messages: str | list[LLMMessage] | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> None:
         """Emit LLM call completed event."""
         from crewai.utilities.serialization import to_serializable
@@ -426,6 +475,7 @@ class BaseLLM(ABC):
                 from_agent=from_agent,
                 model=self.model,
                 call_id=get_current_call_id(),
+                usage=usage,
             ),
         )
 

lib/crewai/src/crewai/llms/providers/anthropic/completion.py

@@ -3,12 +3,13 @@ from __future__ import annotations
 import json
 import logging
 import os
-from typing import TYPE_CHECKING, Any, Final, Literal, TypeGuard, cast
+from typing import Any, Final, Literal, TypeGuard, cast
 
-from pydantic import BaseModel
+from pydantic import BaseModel, PrivateAttr, model_validator
 
 from crewai.events.types.llm_events import LLMCallType
-from crewai.llms.base_llm import BaseLLM, llm_call_context
+from crewai.llms.base_llm import BaseLLM, JsonResponseFormat, llm_call_context
+from crewai.llms.hooks.base import BaseInterceptor
 from crewai.llms.hooks.transport import AsyncHTTPTransport, HTTPTransport
 from crewai.utilities.agent_utils import is_context_length_exceeded
 from crewai.utilities.exceptions.context_window_exceeding_exception import (
@@ -17,9 +18,6 @@ from crewai.utilities.exceptions.context_window_exceeding_exception import (
 from crewai.utilities.types import LLMMessage
 
 
-if TYPE_CHECKING:
-    from crewai.llms.hooks.base import BaseInterceptor
-
 try:
     from anthropic import Anthropic, AsyncAnthropic, transform_schema
     from anthropic.types import (
@@ -150,60 +148,47 @@ class AnthropicCompletion(BaseLLM):
     offering native tool use, streaming support, and proper message formatting.
     """
 
-    def __init__(
-        self,
-        model: str = "claude-3-5-sonnet-20241022",
-        api_key: str | None = None,
-        base_url: str | None = None,
-        timeout: float | None = None,
-        max_retries: int = 2,
-        temperature: float | None = None,
-        max_tokens: int = 4096,  # Required for Anthropic
-        top_p: float | None = None,
-        stop_sequences: list[str] | None = None,
-        stream: bool = False,
-        client_params: dict[str, Any] | None = None,
-        interceptor: BaseInterceptor[httpx.Request, httpx.Response] | None = None,
-        thinking: AnthropicThinkingConfig | None = None,
-        response_format: type[BaseModel] | None = None,
-        tool_search: AnthropicToolSearchConfig | bool | None = None,
-        **kwargs: Any,
-    ):
-        """Initialize Anthropic chat completion client.
+    model: str = "claude-3-5-sonnet-20241022"
+    timeout: float | None = None
+    max_retries: int = 2
+    max_tokens: int = 4096
+    top_p: float | None = None
+    stream: bool = False
+    client_params: dict[str, Any] | None = None
+    interceptor: BaseInterceptor[httpx.Request, httpx.Response] | None = None
+    thinking: AnthropicThinkingConfig | None = None
+    response_format: JsonResponseFormat | type[BaseModel] | None = None
+    tool_search: AnthropicToolSearchConfig | None = None
+    is_claude_3: bool = False
+    supports_tools: bool = True
 
-        Args:
-            model: Anthropic model name (e.g., 'claude-3-5-sonnet-20241022')
-            api_key: Anthropic API key (defaults to ANTHROPIC_API_KEY env var)
-            base_url: Custom base URL for Anthropic API
-            timeout: Request timeout in seconds
-            max_retries: Maximum number of retries
-            temperature: Sampling temperature (0-1)
-            max_tokens: Maximum tokens in response (required for Anthropic)
-            top_p: Nucleus sampling parameter
-            stop_sequences: Stop sequences (Anthropic uses stop_sequences, not stop)
-            stream: Enable streaming responses
-            client_params: Additional parameters for the Anthropic client
-            interceptor: HTTP interceptor for modifying requests/responses at transport level.
-            response_format: Pydantic model for structured output. When provided, responses
-                will be validated against this model schema.
-            tool_search: Enable Anthropic's server-side tool search. When True, uses "bm25"
-                variant by default. Pass an AnthropicToolSearchConfig to choose "regex" or
-                "bm25". When enabled, tools are automatically marked with defer_loading=True
-                and a tool search tool is injected into the tools list.
-            **kwargs: Additional parameters
-        """
-        super().__init__(
-            model=model, temperature=temperature, stop=stop_sequences or [], **kwargs
-        )
+    _client: Any = PrivateAttr(default=None)
+    _async_client: Any = PrivateAttr(default=None)
+    _previous_thinking_blocks: list[Any] = PrivateAttr(default_factory=list)
 
-        # Client params
-        self.interceptor = interceptor
-        self.client_params = client_params
-        self.base_url = base_url
-        self.timeout = timeout
-        self.max_retries = max_retries
+    @model_validator(mode="before")
+    @classmethod
+    def _normalize_anthropic_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+        # Anthropic uses stop_sequences; normalize from stop kwarg
+        popped = data.pop("stop_sequences", None)
+        seqs = popped if popped is not None else (data.get("stop") or [])
+        if isinstance(seqs, str):
+            seqs = [seqs]
+        data["stop"] = seqs
+        data["is_claude_3"] = "claude-3" in data.get("model", "").lower()
+        # Normalize tool_search
+        ts = data.get("tool_search")
+        if ts is True:
+            data["tool_search"] = AnthropicToolSearchConfig()
+        elif ts is not None and not isinstance(ts, AnthropicToolSearchConfig):
+            data["tool_search"] = None
+        return data
 
-        self.client = Anthropic(**self._get_client_params())
+    @model_validator(mode="after")
+    def _init_clients(self) -> AnthropicCompletion:
+        self._client = Anthropic(**self._get_client_params())
 
         async_client_params = self._get_client_params()
         if self.interceptor:
@@ -211,51 +196,8 @@ class AnthropicCompletion(BaseLLM):
             async_http_client = httpx.AsyncClient(transport=async_transport)
             async_client_params["http_client"] = async_http_client
 
-        self.async_client = AsyncAnthropic(**async_client_params)
-
-        # Store completion parameters
-        self.max_tokens = max_tokens
-        self.top_p = top_p
-        self.stream = stream
-        self.stop_sequences = stop_sequences or []
-        self.thinking = thinking
-        self.previous_thinking_blocks: list[ThinkingBlock] = []
-        self.response_format = response_format
-        # Tool search config
-        self.tool_search: AnthropicToolSearchConfig | None
-        if tool_search is True:
-            self.tool_search = AnthropicToolSearchConfig()
-        elif isinstance(tool_search, AnthropicToolSearchConfig):
-            self.tool_search = tool_search
-        else:
-            self.tool_search = None
-        # Model-specific settings
-        self.is_claude_3 = "claude-3" in model.lower()
-        self.supports_tools = True
-
-    @property
-    def stop(self) -> list[str]:
-        """Get stop sequences sent to the API."""
-        return self.stop_sequences
-
-    @stop.setter
-    def stop(self, value: list[str] | str | None) -> None:
-        """Set stop sequences.
-
-        Synchronizes stop_sequences to ensure values set by CrewAgentExecutor
-        are properly sent to the Anthropic API.
-
-        Args:
-            value: Stop sequences as a list, single string, or None
-        """
-        if value is None:
-            self.stop_sequences = []
-        elif isinstance(value, str):
-            self.stop_sequences = [value]
-        elif isinstance(value, list):
-            self.stop_sequences = value
-        else:
-            self.stop_sequences = []
+        self._async_client = AsyncAnthropic(**async_client_params)
+        return self
 
     def to_config_dict(self) -> dict[str, Any]:
         """Extend base config with Anthropic-specific fields."""
@@ -751,11 +693,11 @@ class AnthropicCompletion(BaseLLM):
                         )
                 elif isinstance(content, list):
                     formatted_messages.append({"role": "assistant", "content": content})
-                elif self.thinking and self.previous_thinking_blocks:
+                elif self.thinking and self._previous_thinking_blocks:
                     structured_content = cast(
                         list[dict[str, Any]],
                         [
-                            *self.previous_thinking_blocks,
+                            *self._previous_thinking_blocks,
                             {"type": "text", "text": content if content else ""},
                         ],
                     )
@@ -809,7 +751,7 @@ class AnthropicCompletion(BaseLLM):
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
         from_agent: Any | None = None,
-        response_model: type[BaseModel] | None = None,
+        response_model: JsonResponseFormat | type[BaseModel] | None = None,
     ) -> str | Any:
         """Handle non-streaming message completion."""
         uses_file_api = _contains_file_id_reference(params.get("messages", []))
@@ -843,11 +785,11 @@ class AnthropicCompletion(BaseLLM):
         try:
             if betas:
                 params["betas"] = betas
-                response = self.client.beta.messages.create(
+                response = self._client.beta.messages.create(
                     **params, extra_body=extra_body
                 )
             else:
-                response = self.client.messages.create(**params)
+                response = self._client.messages.create(**params)
 
         except Exception as e:
             if is_context_length_exceeded(e):
@@ -869,6 +811,7 @@ class AnthropicCompletion(BaseLLM):
                             from_task=from_task,
                             from_agent=from_agent,
                             messages=params["messages"],
+                            usage=usage,
                         )
                         return structured_data
             else:
@@ -884,6 +827,7 @@ class AnthropicCompletion(BaseLLM):
                             from_task=from_task,
                             from_agent=from_agent,
                             messages=params["messages"],
+                            usage=usage,
                         )
                         return structured_data
 
@@ -906,6 +850,7 @@ class AnthropicCompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return list(tool_uses)
 
@@ -928,7 +873,7 @@ class AnthropicCompletion(BaseLLM):
                         thinking_blocks.append(cast(ThinkingBlock, thinking_block))
 
         if thinking_blocks:
-            self.previous_thinking_blocks = thinking_blocks
+            self._previous_thinking_blocks = thinking_blocks
 
         content = self._apply_stop_words(content)
         self._emit_call_completed_event(
@@ -937,6 +882,7 @@ class AnthropicCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage,
         )
 
         if usage.get("total_tokens", 0) > 0:
@@ -952,7 +898,7 @@ class AnthropicCompletion(BaseLLM):
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
         from_agent: Any | None = None,
-        response_model: type[BaseModel] | None = None,
+        response_model: JsonResponseFormat | type[BaseModel] | None = None,
     ) -> str | Any:
         """Handle streaming message completion."""
         betas: list[str] = []
@@ -991,9 +937,9 @@ class AnthropicCompletion(BaseLLM):
         current_tool_calls: dict[int, dict[str, Any]] = {}
 
         stream_context = (
-            self.client.beta.messages.stream(**stream_params, extra_body=extra_body)
+            self._client.beta.messages.stream(**stream_params, extra_body=extra_body)
             if betas
-            else self.client.messages.stream(**stream_params)
+            else self._client.messages.stream(**stream_params)
         )
         with stream_context as stream:
             response_id = None
@@ -1072,7 +1018,7 @@ class AnthropicCompletion(BaseLLM):
                     thinking_blocks.append(cast(ThinkingBlock, thinking_block))
 
         if thinking_blocks:
-            self.previous_thinking_blocks = thinking_blocks
+            self._previous_thinking_blocks = thinking_blocks
 
         usage = self._extract_anthropic_token_usage(final_message)
         self._track_token_usage_internal(usage)
@@ -1086,6 +1032,7 @@ class AnthropicCompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage,
                 )
                 return structured_data
             for block in final_message.content:
@@ -1100,6 +1047,7 @@ class AnthropicCompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return structured_data
 
@@ -1129,6 +1077,7 @@ class AnthropicCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage,
         )
 
         return self._invoke_after_llm_call_hooks(
@@ -1269,7 +1218,7 @@ class AnthropicCompletion(BaseLLM):
 
         try:
             # Send tool results back to Claude for final response
-            final_response: Message = self.client.messages.create(**follow_up_params)
+            final_response: Message = self._client.messages.create(**follow_up_params)
 
             # Track token usage for follow-up call
             follow_up_usage = self._extract_anthropic_token_usage(final_response)
@@ -1288,7 +1237,7 @@ class AnthropicCompletion(BaseLLM):
                             thinking_blocks.append(cast(ThinkingBlock, thinking_block))
 
             if thinking_blocks:
-                self.previous_thinking_blocks = thinking_blocks
+                self._previous_thinking_blocks = thinking_blocks
 
             final_content = self._apply_stop_words(final_content)
 
@@ -1299,6 +1248,7 @@ class AnthropicCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=follow_up_params["messages"],
+                usage=follow_up_usage,
             )
 
             # Log combined token usage
@@ -1330,7 +1280,7 @@ class AnthropicCompletion(BaseLLM):
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
         from_agent: Any | None = None,
-        response_model: type[BaseModel] | None = None,
+        response_model: JsonResponseFormat | type[BaseModel] | None = None,
     ) -> str | Any:
         """Handle non-streaming async message completion."""
         uses_file_api = _contains_file_id_reference(params.get("messages", []))
@@ -1364,11 +1314,11 @@ class AnthropicCompletion(BaseLLM):
         try:
             if betas:
                 params["betas"] = betas
-                response = await self.async_client.beta.messages.create(
+                response = await self._async_client.beta.messages.create(
                     **params, extra_body=extra_body
                 )
             else:
-                response = await self.async_client.messages.create(**params)
+                response = await self._async_client.messages.create(**params)
 
         except Exception as e:
             if is_context_length_exceeded(e):
@@ -1390,6 +1340,7 @@ class AnthropicCompletion(BaseLLM):
                             from_task=from_task,
                             from_agent=from_agent,
                             messages=params["messages"],
+                            usage=usage,
                         )
                         return structured_data
             else:
@@ -1405,6 +1356,7 @@ class AnthropicCompletion(BaseLLM):
                             from_task=from_task,
                             from_agent=from_agent,
                             messages=params["messages"],
+                            usage=usage,
                         )
                         return structured_data
 
@@ -1425,6 +1377,7 @@ class AnthropicCompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return list(tool_uses)
 
@@ -1448,6 +1401,7 @@ class AnthropicCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage,
         )
 
         if usage.get("total_tokens", 0) > 0:
@@ -1461,7 +1415,7 @@ class AnthropicCompletion(BaseLLM):
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
         from_agent: Any | None = None,
-        response_model: type[BaseModel] | None = None,
+        response_model: JsonResponseFormat | type[BaseModel] | None = None,
     ) -> str | Any:
         """Handle async streaming message completion."""
         betas: list[str] = []
@@ -1498,11 +1452,11 @@ class AnthropicCompletion(BaseLLM):
         current_tool_calls: dict[int, dict[str, Any]] = {}
 
         stream_context = (
-            self.async_client.beta.messages.stream(
+            self._async_client.beta.messages.stream(
                 **stream_params, extra_body=extra_body
             )
             if betas
-            else self.async_client.messages.stream(**stream_params)
+            else self._async_client.messages.stream(**stream_params)
         )
         async with stream_context as stream:
             response_id = None
@@ -1585,6 +1539,7 @@ class AnthropicCompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage,
                 )
                 return structured_data
             for block in final_message.content:
@@ -1599,6 +1554,7 @@ class AnthropicCompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return structured_data
 
@@ -1627,6 +1583,7 @@ class AnthropicCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage,
         )
 
         return full_response
@@ -1664,7 +1621,7 @@ class AnthropicCompletion(BaseLLM):
         ]
 
         try:
-            final_response: Message = await self.async_client.messages.create(
+            final_response: Message = await self._async_client.messages.create(
                 **follow_up_params
             )
 
@@ -1685,6 +1642,7 @@ class AnthropicCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=follow_up_params["messages"],
+                usage=follow_up_usage,
             )
 
             total_usage = {
@@ -1786,8 +1744,8 @@ class AnthropicCompletion(BaseLLM):
             from crewai_files.uploaders.anthropic import AnthropicFileUploader
 
             return AnthropicFileUploader(
-                client=self.client,
-                async_client=self.async_client,
+                client=self._client,
+                async_client=self._async_client,
             )
         except ImportError:
             return None

lib/crewai/src/crewai/llms/providers/azure/completion.py

@@ -3,11 +3,13 @@ from __future__ import annotations
 import json
 import logging
 import os
-from typing import TYPE_CHECKING, Any, TypedDict
+from typing import Any, TypedDict
+from urllib.parse import urlparse
 
-from pydantic import BaseModel
+from pydantic import BaseModel, PrivateAttr, model_validator
 from typing_extensions import Self
 
+from crewai.llms.hooks.base import BaseInterceptor
 from crewai.utilities.agent_utils import is_context_length_exceeded
 from crewai.utilities.exceptions.context_window_exceeding_exception import (
     LLMContextLengthExceededError,
@@ -16,10 +18,6 @@ from crewai.utilities.pydantic_schema_utils import generate_model_description
 from crewai.utilities.types import LLMMessage
 
 
-if TYPE_CHECKING:
-    from crewai.llms.hooks.base import BaseInterceptor
-
-
 try:
     from azure.ai.inference import (
         ChatCompletionsClient,
@@ -76,109 +74,84 @@ class AzureCompletion(BaseLLM):
     offering native function calling, streaming support, and proper Azure authentication.
     """
 
-    def __init__(
-        self,
-        model: str,
-        api_key: str | None = None,
-        endpoint: str | None = None,
-        api_version: str | None = None,
-        timeout: float | None = None,
-        max_retries: int = 2,
-        temperature: float | None = None,
-        top_p: float | None = None,
-        frequency_penalty: float | None = None,
-        presence_penalty: float | None = None,
-        max_tokens: int | None = None,
-        stop: list[str] | None = None,
-        stream: bool = False,
-        interceptor: BaseInterceptor[Any, Any] | None = None,
-        response_format: type[BaseModel] | None = None,
-        **kwargs: Any,
-    ):
-        """Initialize Azure AI Inference chat completion client.
+    endpoint: str | None = None
+    api_version: str | None = None
+    timeout: float | None = None
+    max_retries: int = 2
+    top_p: float | None = None
+    frequency_penalty: float | None = None
+    presence_penalty: float | None = None
+    max_tokens: int | None = None
+    stream: bool = False
+    interceptor: BaseInterceptor[Any, Any] | None = None
+    response_format: type[BaseModel] | None = None
+    is_openai_model: bool = False
+    is_azure_openai_endpoint: bool = False
 
-        Args:
-            model: Azure deployment name or model name
-            api_key: Azure API key (defaults to AZURE_API_KEY env var)
-            endpoint: Azure endpoint URL (defaults to AZURE_ENDPOINT env var)
-            api_version: Azure API version (defaults to AZURE_API_VERSION env var)
-            timeout: Request timeout in seconds
-            max_retries: Maximum number of retries
-            temperature: Sampling temperature (0-2)
-            top_p: Nucleus sampling parameter
-            frequency_penalty: Frequency penalty (-2 to 2)
-            presence_penalty: Presence penalty (-2 to 2)
-            max_tokens: Maximum tokens in response
-            stop: Stop sequences
-            stream: Enable streaming responses
-            interceptor: HTTP interceptor (not yet supported for Azure).
-            response_format: Pydantic model for structured output. Used as default when
-                           response_model is not passed to call()/acall() methods.
-                           Only works with OpenAI models deployed on Azure.
-            **kwargs: Additional parameters
-        """
-        if interceptor is not None:
+    _client: Any = PrivateAttr(default=None)
+    _async_client: Any = PrivateAttr(default=None)
+
+    @model_validator(mode="before")
+    @classmethod
+    def _normalize_azure_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+
+        if data.get("interceptor") is not None:
             raise NotImplementedError(
                 "HTTP interceptors are not yet supported for Azure AI Inference provider. "
                 "Interceptors are currently supported for OpenAI and Anthropic providers only."
             )
 
-        super().__init__(
-            model=model, temperature=temperature, stop=stop or [], **kwargs
-        )
-
-        self.api_key = api_key or os.getenv("AZURE_API_KEY")
-        self.endpoint = (
-            endpoint
+        # Resolve env vars
+        data["api_key"] = data.get("api_key") or os.getenv("AZURE_API_KEY")
+        data["endpoint"] = (
+            data.get("endpoint")
             or os.getenv("AZURE_ENDPOINT")
             or os.getenv("AZURE_OPENAI_ENDPOINT")
             or os.getenv("AZURE_API_BASE")
         )
-        self.api_version = api_version or os.getenv("AZURE_API_VERSION") or "2024-06-01"
-        self.timeout = timeout
-        self.max_retries = max_retries
+        data["api_version"] = (
+            data.get("api_version") or os.getenv("AZURE_API_VERSION") or "2024-06-01"
+        )
 
-        if not self.api_key:
+        if not data["api_key"]:
             raise ValueError(
                 "Azure API key is required. Set AZURE_API_KEY environment variable or pass api_key parameter."
             )
-        if not self.endpoint:
+        if not data["endpoint"]:
             raise ValueError(
                 "Azure endpoint is required. Set AZURE_ENDPOINT environment variable or pass endpoint parameter."
             )
 
-        # Validate and potentially fix Azure OpenAI endpoint URL
-        self.endpoint = self._validate_and_fix_endpoint(self.endpoint, model)
+        model = data.get("model", "")
+        data["endpoint"] = AzureCompletion._validate_and_fix_endpoint(
+            data["endpoint"], model
+        )
+        data["is_openai_model"] = any(
+            prefix in model.lower() for prefix in ["gpt-", "o1-", "text-"]
+        )
+        parsed = urlparse(data["endpoint"])
+        hostname = parsed.hostname or ""
+        data["is_azure_openai_endpoint"] = (
+            hostname == "openai.azure.com" or hostname.endswith(".openai.azure.com")
+        ) and "/openai/deployments/" in data["endpoint"]
+        return data
 
-        # Build client kwargs
-        client_kwargs = {
+    @model_validator(mode="after")
+    def _init_clients(self) -> AzureCompletion:
+        if not self.api_key:
+            raise ValueError("Azure API key is required.")
+        client_kwargs: dict[str, Any] = {
             "endpoint": self.endpoint,
             "credential": AzureKeyCredential(self.api_key),
         }
-
-        # Add api_version if specified (primarily for Azure OpenAI endpoints)
         if self.api_version:
             client_kwargs["api_version"] = self.api_version
 
-        self.client = ChatCompletionsClient(**client_kwargs)  # type: ignore[arg-type]
-
-        self.async_client = AsyncChatCompletionsClient(**client_kwargs)  # type: ignore[arg-type]
-
-        self.top_p = top_p
-        self.frequency_penalty = frequency_penalty
-        self.presence_penalty = presence_penalty
-        self.max_tokens = max_tokens
-        self.stream = stream
-        self.response_format = response_format
-
-        self.is_openai_model = any(
-            prefix in model.lower() for prefix in ["gpt-", "o1-", "text-"]
-        )
-
-        self.is_azure_openai_endpoint = (
-            "openai.azure.com" in self.endpoint
-            and "/openai/deployments/" in self.endpoint
-        )
+        self._client = ChatCompletionsClient(**client_kwargs)
+        self._async_client = AsyncChatCompletionsClient(**client_kwargs)
+        return self
 
     def to_config_dict(self) -> dict[str, Any]:
         """Extend base config with Azure-specific fields."""
@@ -215,7 +188,11 @@ class AzureCompletion(BaseLLM):
         Returns:
             Validated and potentially corrected endpoint URL
         """
-        if "openai.azure.com" in endpoint and "/openai/deployments/" not in endpoint:
+        ep_host = urlparse(endpoint).hostname or ""
+        is_azure_openai = ep_host == "openai.azure.com" or ep_host.endswith(
+            ".openai.azure.com"
+        )
+        if is_azure_openai and "/openai/deployments/" not in endpoint:
             endpoint = endpoint.rstrip("/")
 
             if not endpoint.endswith("/openai/deployments"):
@@ -592,6 +569,7 @@ class AzureCompletion(BaseLLM):
         params: AzureCompletionParams,
         from_task: Any | None = None,
         from_agent: Any | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> BaseModel:
         """Validate content against response model and emit completion event.
 
@@ -617,6 +595,7 @@ class AzureCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage,
             )
 
             return structured_data
@@ -666,6 +645,7 @@ class AzureCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage,
             )
             return list(message.tool_calls)
 
@@ -703,6 +683,7 @@ class AzureCompletion(BaseLLM):
                 params=params,
                 from_task=from_task,
                 from_agent=from_agent,
+                usage=usage,
             )
 
         content = self._apply_stop_words(content)
@@ -714,6 +695,7 @@ class AzureCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage,
         )
 
         return self._invoke_after_llm_call_hooks(
@@ -731,7 +713,7 @@ class AzureCompletion(BaseLLM):
         """Handle non-streaming chat completion."""
         try:
             # Cast params to Any to avoid type checking issues with TypedDict unpacking
-            response: ChatCompletions = self.client.complete(**params)  # type: ignore[assignment,arg-type]
+            response: ChatCompletions = self._client.complete(**params)
             return self._process_completion_response(
                 response=response,
                 params=params,
@@ -817,7 +799,7 @@ class AzureCompletion(BaseLLM):
         self,
         full_response: str,
         tool_calls: dict[int, dict[str, Any]],
-        usage_data: dict[str, int],
+        usage_data: dict[str, Any] | None,
         params: AzureCompletionParams,
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
@@ -829,7 +811,7 @@ class AzureCompletion(BaseLLM):
         Args:
             full_response: The complete streamed response content
             tool_calls: Dictionary of tool calls accumulated during streaming
-            usage_data: Token usage data from the stream
+            usage_data: Token usage data from the stream, or None if unavailable
             params: Completion parameters containing messages
             available_functions: Available functions for tool calling
             from_task: Task that initiated the call
@@ -839,7 +821,8 @@ class AzureCompletion(BaseLLM):
         Returns:
             Final response content after processing, or structured output
         """
-        self._track_token_usage_internal(usage_data)
+        if usage_data:
+            self._track_token_usage_internal(usage_data)
 
         # Handle structured output validation
         if response_model and self.is_openai_model:
@@ -849,6 +832,7 @@ class AzureCompletion(BaseLLM):
                 params=params,
                 from_task=from_task,
                 from_agent=from_agent,
+                usage=usage_data,
             )
 
         # If there are tool_calls but no available_functions, return them
@@ -871,6 +855,7 @@ class AzureCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage_data,
             )
             return formatted_tool_calls
 
@@ -907,6 +892,7 @@ class AzureCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage_data,
         )
 
         return self._invoke_after_llm_call_hooks(
@@ -925,8 +911,8 @@ class AzureCompletion(BaseLLM):
         full_response = ""
         tool_calls: dict[int, dict[str, Any]] = {}
 
-        usage_data = {"total_tokens": 0}
-        for update in self.client.complete(**params):  # type: ignore[arg-type]
+        usage_data: dict[str, Any] | None = None
+        for update in self._client.complete(**params):
             if isinstance(update, StreamingChatCompletionsUpdate):
                 if update.usage:
                     usage = update.usage
@@ -967,7 +953,7 @@ class AzureCompletion(BaseLLM):
         """Handle non-streaming chat completion asynchronously."""
         try:
             # Cast params to Any to avoid type checking issues with TypedDict unpacking
-            response: ChatCompletions = await self.async_client.complete(**params)  # type: ignore[assignment,arg-type]
+            response: ChatCompletions = await self._async_client.complete(**params)
             return self._process_completion_response(
                 response=response,
                 params=params,
@@ -991,10 +977,10 @@ class AzureCompletion(BaseLLM):
         full_response = ""
         tool_calls: dict[int, dict[str, Any]] = {}
 
-        usage_data = {"total_tokens": 0}
+        usage_data: dict[str, Any] | None = None
 
-        stream = await self.async_client.complete(**params)  # type: ignore[arg-type]
-        async for update in stream:  # type: ignore[union-attr]
+        stream = await self._async_client.complete(**params)
+        async for update in stream:
             if isinstance(update, StreamingChatCompletionsUpdate):
                 if hasattr(update, "usage") and update.usage:
                     usage = update.usage
@@ -1110,8 +1096,8 @@ class AzureCompletion(BaseLLM):
         This ensures proper cleanup of the underlying aiohttp session
         to avoid unclosed connector warnings.
         """
-        if hasattr(self.async_client, "close"):
-            await self.async_client.close()
+        if hasattr(self._async_client, "close"):
+            await self._async_client.close()
 
     async def __aenter__(self) -> Self:
         """Async context manager entry."""

lib/crewai/src/crewai/llms/providers/bedrock/completion.py

@@ -7,7 +7,7 @@ import logging
 import os
 from typing import TYPE_CHECKING, Any, TypedDict, cast
 
-from pydantic import BaseModel
+from pydantic import BaseModel, PrivateAttr, model_validator
 from typing_extensions import Required
 
 from crewai.events.types.llm_events import LLMCallType
@@ -33,7 +33,7 @@ if TYPE_CHECKING:
         ToolTypeDef,
     )
 
-    from crewai.llms.hooks.base import BaseInterceptor
+from crewai.llms.hooks.base import BaseInterceptor
 
 
 try:
@@ -228,129 +228,97 @@ class BedrockCompletion(BaseLLM):
     - Model-specific conversation format handling (e.g., Cohere requirements)
     """
 
-    def __init__(
-        self,
-        model: str = "anthropic.claude-3-5-sonnet-20241022-v2:0",
-        aws_access_key_id: str | None = None,
-        aws_secret_access_key: str | None = None,
-        aws_session_token: str | None = None,
-        region_name: str | None = None,
-        temperature: float | None = None,
-        max_tokens: int | None = None,
-        top_p: float | None = None,
-        top_k: int | None = None,
-        stop_sequences: Sequence[str] | None = None,
-        stream: bool = False,
-        guardrail_config: dict[str, Any] | None = None,
-        additional_model_request_fields: dict[str, Any] | None = None,
-        additional_model_response_field_paths: list[str] | None = None,
-        interceptor: BaseInterceptor[Any, Any] | None = None,
-        response_format: type[BaseModel] | None = None,
-        **kwargs: Any,
-    ) -> None:
-        """Initialize AWS Bedrock completion client.
+    model: str = "anthropic.claude-3-5-sonnet-20241022-v2:0"
+    aws_access_key_id: str | None = None
+    aws_secret_access_key: str | None = None
+    aws_session_token: str | None = None
+    region_name: str | None = None
+    max_tokens: int | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    stream: bool = False
+    guardrail_config: dict[str, Any] | None = None
+    additional_model_request_fields: dict[str, Any] | None = None
+    additional_model_response_field_paths: list[str] | None = None
+    interceptor: BaseInterceptor[Any, Any] | None = None
+    response_format: type[BaseModel] | None = None
+    is_claude_model: bool = False
+    supports_tools: bool = True
+    supports_streaming: bool = True
+    model_id: str = ""
 
-        Args:
-            model: The Bedrock model ID to use
-            aws_access_key_id: AWS access key (defaults to environment variable)
-            aws_secret_access_key: AWS secret key (defaults to environment variable)
-            aws_session_token: AWS session token for temporary credentials
-            region_name: AWS region name
-            temperature: Sampling temperature for response generation
-            max_tokens: Maximum tokens to generate
-            top_p: Nucleus sampling parameter
-            top_k: Top-k sampling parameter (Claude models only)
-            stop_sequences: List of sequences that stop generation
-            stream: Whether to use streaming responses
-            guardrail_config: Guardrail configuration for content filtering
-            additional_model_request_fields: Model-specific request parameters
-            additional_model_response_field_paths: Custom response field paths
-            interceptor: HTTP interceptor (not yet supported for Bedrock).
-            response_format: Pydantic model for structured output. Used as default when
-                           response_model is not passed to call()/acall() methods.
-            **kwargs: Additional parameters
-        """
-        if interceptor is not None:
+    _client: Any = PrivateAttr(default=None)
+    _async_exit_stack: Any = PrivateAttr(default=None)
+    _async_client_initialized: bool = PrivateAttr(default=False)
+    _async_client: Any = PrivateAttr(default=None)
+
+    @model_validator(mode="before")
+    @classmethod
+    def _normalize_bedrock_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+
+        if data.get("interceptor") is not None:
             raise NotImplementedError(
                 "HTTP interceptors are not yet supported for AWS Bedrock provider. "
                 "Interceptors are currently supported for OpenAI and Anthropic providers only."
             )
 
-        # Extract provider from kwargs to avoid duplicate argument
-        kwargs.pop("provider", None)
+        # Force provider to bedrock
+        data.pop("provider", None)
+        data["provider"] = "bedrock"
 
-        super().__init__(
-            model=model,
-            temperature=temperature,
-            stop=stop_sequences or [],
-            provider="bedrock",
-            **kwargs,
+        # Normalize stop_sequences from stop kwarg
+        popped = data.pop("stop_sequences", None)
+        seqs = popped if popped is not None else (data.get("stop") or [])
+        if isinstance(seqs, str):
+            seqs = [seqs]
+        elif isinstance(seqs, Sequence) and not isinstance(seqs, list):
+            seqs = list(seqs)
+        data["stop"] = seqs
+
+        # Resolve env vars
+        data["aws_access_key_id"] = data.get("aws_access_key_id") or os.getenv(
+            "AWS_ACCESS_KEY_ID"
         )
-
-        # Configure client with timeouts and retries following AWS best practices
-        config = Config(
-            read_timeout=300,
-            retries={
-                "max_attempts": 3,
-                "mode": "adaptive",
-            },
-            tcp_keepalive=True,
+        data["aws_secret_access_key"] = data.get("aws_secret_access_key") or os.getenv(
+            "AWS_SECRET_ACCESS_KEY"
         )
-
-        self.region_name = (
-            region_name
+        data["aws_session_token"] = data.get("aws_session_token") or os.getenv(
+            "AWS_SESSION_TOKEN"
+        )
+        data["region_name"] = (
+            data.get("region_name")
             or os.getenv("AWS_DEFAULT_REGION")
             or os.getenv("AWS_REGION_NAME")
             or "us-east-1"
         )
 
-        self.aws_access_key_id = aws_access_key_id or os.getenv("AWS_ACCESS_KEY_ID")
-        self.aws_secret_access_key = aws_secret_access_key or os.getenv(
-            "AWS_SECRET_ACCESS_KEY"
-        )
-        self.aws_session_token = aws_session_token or os.getenv("AWS_SESSION_TOKEN")
+        model = data.get("model", "anthropic.claude-3-5-sonnet-20241022-v2:0")
+        data["is_claude_model"] = "claude" in model.lower()
+        data["model_id"] = model
+        return data
 
-        # Initialize Bedrock client with proper configuration
+    @model_validator(mode="after")
+    def _init_clients(self) -> BedrockCompletion:
+        config = Config(
+            read_timeout=300,
+            retries={"max_attempts": 3, "mode": "adaptive"},
+            tcp_keepalive=True,
+        )
         session = Session(
             aws_access_key_id=self.aws_access_key_id,
             aws_secret_access_key=self.aws_secret_access_key,
             aws_session_token=self.aws_session_token,
             region_name=self.region_name,
         )
-
-        self.client = session.client("bedrock-runtime", config=config)
-
+        self._client = session.client("bedrock-runtime", config=config)
         self._async_exit_stack = AsyncExitStack() if AIOBOTOCORE_AVAILABLE else None
-        self._async_client_initialized = False
-
-        # Store completion parameters
-        self.max_tokens = max_tokens
-        self.top_p = top_p
-        self.top_k = top_k
-        self.stream = stream
-        self.stop_sequences = stop_sequences
-        self.response_format = response_format
-
-        # Store advanced features (optional)
-        self.guardrail_config = guardrail_config
-        self.additional_model_request_fields = additional_model_request_fields
-        self.additional_model_response_field_paths = (
-            additional_model_response_field_paths
-        )
-
-        # Model-specific settings
-        self.is_claude_model = "claude" in model.lower()
-        self.supports_tools = True  # Converse API supports tools for most models
-        self.supports_streaming = True
-
-        # Handle inference profiles for newer models
-        self.model_id = model
+        return self
 
     def to_config_dict(self) -> dict[str, Any]:
         """Extend base config with Bedrock-specific fields."""
         config = super().to_config_dict()
-        # NOTE: AWS credentials (access_key, secret_key, session_token) are
-        # intentionally excluded — they must come from env on resume.
         if self.region_name and self.region_name != "us-east-1":
             config["region_name"] = self.region_name
         if self.max_tokens is not None:
@@ -363,30 +331,6 @@ class BedrockCompletion(BaseLLM):
             config["guardrail_config"] = self.guardrail_config
         return config
 
-    @property
-    def stop(self) -> list[str]:
-        """Get stop sequences sent to the API."""
-        return [] if self.stop_sequences is None else list(self.stop_sequences)
-
-    @stop.setter
-    def stop(self, value: Sequence[str] | str | None) -> None:
-        """Set stop sequences.
-
-        Synchronizes stop_sequences to ensure values set by CrewAgentExecutor
-        are properly sent to the Bedrock API.
-
-        Args:
-            value: Stop sequences as a Sequence, single string, or None
-        """
-        if value is None:
-            self.stop_sequences = []
-        elif isinstance(value, str):
-            self.stop_sequences = [value]
-        elif isinstance(value, Sequence):
-            self.stop_sequences = list(value)
-        else:
-            self.stop_sequences = []
-
     def call(
         self,
         messages: str | list[LLMMessage],
@@ -710,7 +654,7 @@ class BedrockCompletion(BaseLLM):
                     raise ValueError(f"Invalid message format at index {i}")
 
             # Call Bedrock Converse API with proper error handling
-            response = self.client.converse(
+            response = self._client.converse(
                 modelId=self.model_id,
                 messages=cast(
                     "Sequence[MessageTypeDef | MessageOutputTypeDef]",
@@ -720,8 +664,9 @@ class BedrockCompletion(BaseLLM):
             )
 
             # Track token usage according to AWS response format
-            if "usage" in response:
-                self._track_token_usage_internal(response["usage"])
+            usage = response.get("usage")
+            if usage:
+                self._track_token_usage_internal(usage)
 
             stop_reason = response.get("stopReason")
             if stop_reason:
@@ -761,6 +706,7 @@ class BedrockCompletion(BaseLLM):
                                 from_task=from_task,
                                 from_agent=from_agent,
                                 messages=messages,
+                                usage=usage,
                             )
                             return result
                         except Exception as e:
@@ -783,6 +729,7 @@ class BedrockCompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=messages,
+                    usage=usage,
                 )
                 return non_structured_output_tool_uses
 
@@ -862,6 +809,7 @@ class BedrockCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=messages,
+                usage=usage,
             )
 
             return self._invoke_after_llm_call_hooks(
@@ -992,15 +940,16 @@ class BedrockCompletion(BaseLLM):
         tool_use_id: str | None = None
         tool_use_index = 0
         accumulated_tool_input = ""
+        usage_data: dict[str, Any] | None = None
 
         try:
-            response = self.client.converse_stream(
+            response = self._client.converse_stream(
                 modelId=self.model_id,
                 messages=cast(
                     "Sequence[MessageTypeDef | MessageOutputTypeDef]",
                     cast(object, messages),
                 ),
-                **body,  # type: ignore[arg-type]
+                **body,
             )
 
             stream = response.get("stream")
@@ -1101,6 +1050,7 @@ class BedrockCompletion(BaseLLM):
                                         from_task=from_task,
                                         from_agent=from_agent,
                                         messages=messages,
+                                        usage=usage_data,
                                     )
                                     return result  # type: ignore[return-value]
                                 except Exception as e:
@@ -1168,6 +1118,7 @@ class BedrockCompletion(BaseLLM):
                         metadata = event["metadata"]
                         if "usage" in metadata:
                             usage_metrics = metadata["usage"]
+                            usage_data = usage_metrics
                             self._track_token_usage_internal(usage_metrics)
                             logging.debug(f"Token usage: {usage_metrics}")
                             if "trace" in metadata:
@@ -1197,6 +1148,7 @@ class BedrockCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=messages,
+            usage=usage_data,
         )
 
         return full_response
@@ -1308,8 +1260,9 @@ class BedrockCompletion(BaseLLM):
                 **body,
             )
 
-            if "usage" in response:
-                self._track_token_usage_internal(response["usage"])
+            usage = response.get("usage")
+            if usage:
+                self._track_token_usage_internal(usage)
 
             stop_reason = response.get("stopReason")
             if stop_reason:
@@ -1348,6 +1301,7 @@ class BedrockCompletion(BaseLLM):
                                 from_task=from_task,
                                 from_agent=from_agent,
                                 messages=messages,
+                                usage=usage,
                             )
                             return result
                         except Exception as e:
@@ -1370,6 +1324,7 @@ class BedrockCompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=messages,
+                    usage=usage,
                 )
                 return non_structured_output_tool_uses
 
@@ -1444,6 +1399,7 @@ class BedrockCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=messages,
+                usage=usage,
             )
 
             return text_content
@@ -1564,6 +1520,7 @@ class BedrockCompletion(BaseLLM):
         tool_use_id: str | None = None
         tool_use_index = 0
         accumulated_tool_input = ""
+        usage_data: dict[str, Any] | None = None
 
         try:
             async_client = await self._ensure_async_client()
@@ -1675,6 +1632,7 @@ class BedrockCompletion(BaseLLM):
                                         from_task=from_task,
                                         from_agent=from_agent,
                                         messages=messages,
+                                        usage=usage_data,
                                     )
                                     return result  # type: ignore[return-value]
                                 except Exception as e:
@@ -1747,6 +1705,7 @@ class BedrockCompletion(BaseLLM):
                         metadata = event["metadata"]
                         if "usage" in metadata:
                             usage_metrics = metadata["usage"]
+                            usage_data = usage_metrics
                             self._track_token_usage_internal(usage_metrics)
                             logging.debug(f"Token usage: {usage_metrics}")
                         if "trace" in metadata:
@@ -1776,6 +1735,7 @@ class BedrockCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=messages,
+            usage=usage_data,
         )
 
         return self._invoke_after_llm_call_hooks(

lib/crewai/src/crewai/llms/providers/gemini/completion.py

@@ -5,12 +5,13 @@ import json
 import logging
 import os
 import re
-from typing import TYPE_CHECKING, Any, Literal, cast
+from typing import Any, Literal, cast
 
-from pydantic import BaseModel
+from pydantic import BaseModel, Field, PrivateAttr, model_validator
 
 from crewai.events.types.llm_events import LLMCallType
 from crewai.llms.base_llm import BaseLLM, llm_call_context
+from crewai.llms.hooks.base import BaseInterceptor
 from crewai.utilities.agent_utils import is_context_length_exceeded
 from crewai.utilities.exceptions.context_window_exceeding_exception import (
     LLMContextLengthExceededError,
@@ -19,10 +20,6 @@ from crewai.utilities.pydantic_schema_utils import generate_model_description
 from crewai.utilities.types import LLMMessage
 
 
-if TYPE_CHECKING:
-    from crewai.llms.hooks.base import BaseInterceptor
-
-
 try:
     from google import genai
     from google.genai import types
@@ -44,137 +41,84 @@ class GeminiCompletion(BaseLLM):
     offering native function calling, streaming support, and proper Gemini formatting.
     """
 
-    def __init__(
-        self,
-        model: str = "gemini-2.0-flash-001",
-        api_key: str | None = None,
-        project: str | None = None,
-        location: str | None = None,
-        temperature: float | None = None,
-        top_p: float | None = None,
-        top_k: int | None = None,
-        max_output_tokens: int | None = None,
-        stop_sequences: list[str] | None = None,
-        stream: bool = False,
-        safety_settings: dict[str, Any] | None = None,
-        client_params: dict[str, Any] | None = None,
-        interceptor: BaseInterceptor[Any, Any] | None = None,
-        use_vertexai: bool | None = None,
-        response_format: type[BaseModel] | None = None,
-        thinking_config: types.ThinkingConfig | None = None,
-        **kwargs: Any,
-    ):
-        """Initialize Google Gemini chat completion client.
+    model: str = "gemini-2.0-flash-001"
+    project: str | None = None
+    location: str | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    max_output_tokens: int | None = None
+    stream: bool = False
+    safety_settings: dict[str, Any] = Field(default_factory=dict)
+    client_params: dict[str, Any] = Field(default_factory=dict)
+    interceptor: BaseInterceptor[Any, Any] | None = None
+    use_vertexai: bool = False
+    response_format: type[BaseModel] | None = None
+    thinking_config: Any = None
+    tools: list[dict[str, Any]] | None = None
+    supports_tools: bool = False
+    is_gemini_2_0: bool = False
 
-        Args:
-            model: Gemini model name (e.g., 'gemini-2.0-flash-001', 'gemini-1.5-pro')
-            api_key: Google API key for Gemini API authentication.
-                    Defaults to GOOGLE_API_KEY or GEMINI_API_KEY env var.
-                    NOTE: Cannot be used with Vertex AI (project parameter). Use Gemini API instead.
-            project: Google Cloud project ID for Vertex AI with ADC authentication.
-                    Requires Application Default Credentials (gcloud auth application-default login).
-                    NOTE: Vertex AI does NOT support API keys, only OAuth2/ADC.
-                    If both api_key and project are set, api_key takes precedence.
-            location: Google Cloud location (for Vertex AI with ADC, defaults to 'us-central1')
-            temperature: Sampling temperature (0-2)
-            top_p: Nucleus sampling parameter
-            top_k: Top-k sampling parameter
-            max_output_tokens: Maximum tokens in response
-            stop_sequences: Stop sequences
-            stream: Enable streaming responses
-            safety_settings: Safety filter settings
-            client_params: Additional parameters to pass to the Google Gen AI Client constructor.
-                          Supports parameters like http_options, credentials, debug_config, etc.
-            interceptor: HTTP interceptor (not yet supported for Gemini).
-            use_vertexai: Whether to use Vertex AI instead of Gemini API.
-                         - True: Use Vertex AI (with ADC or Express mode with API key)
-                         - False: Use Gemini API (explicitly override env var)
-                         - None (default): Check GOOGLE_GENAI_USE_VERTEXAI env var
-                         When using Vertex AI with API key (Express mode), http_options with
-                         api_version="v1" is automatically configured.
-            response_format: Pydantic model for structured output. Used as default when
-                           response_model is not passed to call()/acall() methods.
-            thinking_config: ThinkingConfig for thinking models (gemini-2.5+, gemini-3+).
-                           Controls thought output via include_thoughts, thinking_budget,
-                           and thinking_level. When None, thinking models automatically
-                           get include_thoughts=True so thought content is surfaced.
-            **kwargs: Additional parameters
-        """
-        if interceptor is not None:
+    _client: Any = PrivateAttr(default=None)
+
+    @model_validator(mode="before")
+    @classmethod
+    def _normalize_gemini_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+
+        if data.get("interceptor") is not None:
             raise NotImplementedError(
                 "HTTP interceptors are not yet supported for Google Gemini provider. "
                 "Interceptors are currently supported for OpenAI and Anthropic providers only."
             )
 
-        super().__init__(
-            model=model, temperature=temperature, stop=stop_sequences or [], **kwargs
+        # Normalize stop_sequences from stop kwarg
+        popped = data.pop("stop_sequences", None)
+        seqs = popped if popped is not None else (data.get("stop") or [])
+        if isinstance(seqs, str):
+            seqs = [seqs]
+        data["stop"] = seqs
+
+        # Resolve env vars
+        data["api_key"] = (
+            data.get("api_key")
+            or os.getenv("GOOGLE_API_KEY")
+            or os.getenv("GEMINI_API_KEY")
+        )
+        data["project"] = data.get("project") or os.getenv("GOOGLE_CLOUD_PROJECT")
+        data["location"] = (
+            data.get("location") or os.getenv("GOOGLE_CLOUD_LOCATION") or "us-central1"
         )
 
-        # Store client params for later use
-        self.client_params = client_params or {}
-
-        # Get API configuration with environment variable fallbacks
-        self.api_key = (
-            api_key or os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
-        )
-        self.project = project or os.getenv("GOOGLE_CLOUD_PROJECT")
-        self.location = location or os.getenv("GOOGLE_CLOUD_LOCATION") or "us-central1"
-
-        if use_vertexai is None:
-            use_vertexai = os.getenv("GOOGLE_GENAI_USE_VERTEXAI", "").lower() == "true"
-
-        self.client = self._initialize_client(use_vertexai)
-
-        # Store completion parameters
-        self.top_p = top_p
-        self.top_k = top_k
-        self.max_output_tokens = max_output_tokens
-        self.stream = stream
-        self.safety_settings = safety_settings or {}
-        self.stop_sequences = stop_sequences or []
-        self.tools: list[dict[str, Any]] | None = None
-        self.response_format = response_format
+        use_vx = data.get("use_vertexai")
+        if use_vx is None:
+            use_vx = os.getenv("GOOGLE_GENAI_USE_VERTEXAI", "").lower() == "true"
+        data["use_vertexai"] = use_vx
 
         # Model-specific settings
+        model = data.get("model", "gemini-2.0-flash-001")
         version_match = re.search(r"gemini-(\d+(?:\.\d+)?)", model.lower())
-        self.supports_tools = bool(
+        data["supports_tools"] = bool(
             version_match and float(version_match.group(1)) >= 1.5
         )
-        self.is_gemini_2_0 = bool(
+        data["is_gemini_2_0"] = bool(
             version_match and float(version_match.group(1)) >= 2.0
         )
 
-        self.thinking_config = thinking_config
+        # Auto-enable thinking for gemini-2.5+
         if (
-            self.thinking_config is None
+            data.get("thinking_config") is None
             and version_match
             and float(version_match.group(1)) >= 2.5
         ):
-            self.thinking_config = types.ThinkingConfig(include_thoughts=True)
+            data["thinking_config"] = types.ThinkingConfig(include_thoughts=True)
 
-    @property
-    def stop(self) -> list[str]:
-        """Get stop sequences sent to the API."""
-        return self.stop_sequences
+        return data
 
-    @stop.setter
-    def stop(self, value: list[str] | str | None) -> None:
-        """Set stop sequences.
-
-        Synchronizes stop_sequences to ensure values set by CrewAgentExecutor
-        are properly sent to the Gemini API.
-
-        Args:
-            value: Stop sequences as a list, single string, or None
-        """
-        if value is None:
-            self.stop_sequences = []
-        elif isinstance(value, str):
-            self.stop_sequences = [value]
-        elif isinstance(value, list):
-            self.stop_sequences = value
-        else:
-            self.stop_sequences = []
+    @model_validator(mode="after")
+    def _init_client(self) -> GeminiCompletion:
+        self._client = self._initialize_client(self.use_vertexai)
+        return self
 
     def to_config_dict(self) -> dict[str, Any]:
         """Extend base config with Gemini/Vertex-specific fields."""
@@ -283,8 +227,8 @@ class GeminiCompletion(BaseLLM):
 
         if (
             hasattr(self, "client")
-            and hasattr(self.client, "vertexai")
-            and self.client.vertexai
+            and hasattr(self._client, "vertexai")
+            and self._client.vertexai
         ):
             # Vertex AI configuration
             params.update(
@@ -721,6 +665,7 @@ class GeminiCompletion(BaseLLM):
         messages_for_event: list[LLMMessage],
         from_task: Any | None = None,
         from_agent: Any | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> BaseModel:
         """Validate content against response model and emit completion event.
 
@@ -746,6 +691,7 @@ class GeminiCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=messages_for_event,
+                usage=usage,
             )
 
             return structured_data
@@ -761,6 +707,7 @@ class GeminiCompletion(BaseLLM):
         response_model: type[BaseModel] | None = None,
         from_task: Any | None = None,
         from_agent: Any | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> str | BaseModel:
         """Finalize completion response with validation and event emission.
 
@@ -784,6 +731,7 @@ class GeminiCompletion(BaseLLM):
                 messages_for_event=messages_for_event,
                 from_task=from_task,
                 from_agent=from_agent,
+                usage=usage,
             )
 
         self._emit_call_completed_event(
@@ -792,6 +740,7 @@ class GeminiCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=messages_for_event,
+            usage=usage,
         )
 
         return self._invoke_after_llm_call_hooks(
@@ -805,6 +754,7 @@ class GeminiCompletion(BaseLLM):
         contents: list[types.Content],
         from_task: Any | None = None,
         from_agent: Any | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> BaseModel:
         """Validate and emit event for structured_output tool call.
 
@@ -829,6 +779,7 @@ class GeminiCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=self._convert_contents_to_dict(contents),
+                usage=usage,
             )
             return validated_data
         except Exception as e:
@@ -847,6 +798,7 @@ class GeminiCompletion(BaseLLM):
         from_task: Any | None = None,
         from_agent: Any | None = None,
         response_model: type[BaseModel] | None = None,
+        usage: dict[str, Any] | None = None,
     ) -> str | Any:
         """Process response, execute function calls, and finalize completion.
 
@@ -887,6 +839,7 @@ class GeminiCompletion(BaseLLM):
                                 contents=contents,
                                 from_task=from_task,
                                 from_agent=from_agent,
+                                usage=usage,
                             )
 
                 # Filter out structured_output from function calls returned to executor
@@ -908,6 +861,7 @@ class GeminiCompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=self._convert_contents_to_dict(contents),
+                        usage=usage,
                     )
                     return non_structured_output_parts
 
@@ -949,6 +903,7 @@ class GeminiCompletion(BaseLLM):
             response_model=effective_response_model,
             from_task=from_task,
             from_agent=from_agent,
+            usage=usage,
         )
 
     def _process_stream_chunk(
@@ -956,10 +911,10 @@ class GeminiCompletion(BaseLLM):
         chunk: GenerateContentResponse,
         full_response: str,
         function_calls: dict[int, dict[str, Any]],
-        usage_data: dict[str, int],
+        usage_data: dict[str, int] | None,
         from_task: Any | None = None,
         from_agent: Any | None = None,
-    ) -> tuple[str, dict[int, dict[str, Any]], dict[str, int]]:
+    ) -> tuple[str, dict[int, dict[str, Any]], dict[str, int] | None]:
         """Process a single streaming chunk.
 
         Args:
@@ -1035,7 +990,7 @@ class GeminiCompletion(BaseLLM):
         self,
         full_response: str,
         function_calls: dict[int, dict[str, Any]],
-        usage_data: dict[str, int],
+        usage_data: dict[str, int] | None,
         contents: list[types.Content],
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
@@ -1047,7 +1002,7 @@ class GeminiCompletion(BaseLLM):
         Args:
             full_response: The complete streamed response content
             function_calls: Dictionary of function calls accumulated during streaming
-            usage_data: Token usage data from the stream
+            usage_data: Token usage data from the stream, or None if unavailable
             contents: Original contents for event conversion
             available_functions: Available functions for function calling
             from_task: Task that initiated the call
@@ -1057,7 +1012,8 @@ class GeminiCompletion(BaseLLM):
         Returns:
             Final response content after processing
         """
-        self._track_token_usage_internal(usage_data)
+        if usage_data:
+            self._track_token_usage_internal(usage_data)
 
         if response_model and function_calls:
             for call_data in function_calls.values():
@@ -1069,6 +1025,7 @@ class GeminiCompletion(BaseLLM):
                         contents=contents,
                         from_task=from_task,
                         from_agent=from_agent,
+                        usage=usage_data,
                     )
 
         non_structured_output_calls = {
@@ -1097,6 +1054,7 @@ class GeminiCompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=self._convert_contents_to_dict(contents),
+                usage=usage_data,
             )
             return formatted_function_calls
 
@@ -1137,6 +1095,7 @@ class GeminiCompletion(BaseLLM):
             response_model=effective_response_model,
             from_task=from_task,
             from_agent=from_agent,
+            usage=usage_data,
         )
 
     def _handle_completion(
@@ -1152,7 +1111,7 @@ class GeminiCompletion(BaseLLM):
         try:
             # The API accepts list[Content] but mypy is overly strict about variance
             contents_for_api: Any = contents
-            response = self.client.models.generate_content(
+            response = self._client.models.generate_content(
                 model=self.model,
                 contents=contents_for_api,
                 config=config,
@@ -1174,6 +1133,7 @@ class GeminiCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             response_model=response_model,
+            usage=usage,
         )
 
     def _handle_streaming_completion(
@@ -1188,11 +1148,11 @@ class GeminiCompletion(BaseLLM):
         """Handle streaming content generation."""
         full_response = ""
         function_calls: dict[int, dict[str, Any]] = {}
-        usage_data = {"total_tokens": 0}
+        usage_data: dict[str, int] | None = None
 
         # The API accepts list[Content] but mypy is overly strict about variance
         contents_for_api: Any = contents
-        for chunk in self.client.models.generate_content_stream(
+        for chunk in self._client.models.generate_content_stream(
             model=self.model,
             contents=contents_for_api,
             config=config,
@@ -1230,7 +1190,7 @@ class GeminiCompletion(BaseLLM):
         try:
             # The API accepts list[Content] but mypy is overly strict about variance
             contents_for_api: Any = contents
-            response = await self.client.aio.models.generate_content(
+            response = await self._client.aio.models.generate_content(
                 model=self.model,
                 contents=contents_for_api,
                 config=config,
@@ -1252,6 +1212,7 @@ class GeminiCompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             response_model=response_model,
+            usage=usage,
         )
 
     async def _ahandle_streaming_completion(
@@ -1266,11 +1227,11 @@ class GeminiCompletion(BaseLLM):
         """Handle async streaming content generation."""
         full_response = ""
         function_calls: dict[int, dict[str, Any]] = {}
-        usage_data = {"total_tokens": 0}
+        usage_data: dict[str, int] | None = None
 
         # The API accepts list[Content] but mypy is overly strict about variance
         contents_for_api: Any = contents
-        stream = await self.client.aio.models.generate_content_stream(
+        stream = await self._client.aio.models.generate_content_stream(
             model=self.model,
             contents=contents_for_api,
             config=config,
@@ -1474,6 +1435,6 @@ class GeminiCompletion(BaseLLM):
         try:
             from crewai_files.uploaders.gemini import GeminiFileUploader
 
-            return GeminiFileUploader(client=self.client)
+            return GeminiFileUploader(client=self._client)
         except ImportError:
             return None

lib/crewai/src/crewai/llms/providers/openai/completion.py

@@ -14,10 +14,11 @@ from openai.types.chat import ChatCompletion, ChatCompletionChunk
 from openai.types.chat.chat_completion import Choice
 from openai.types.chat.chat_completion_chunk import ChoiceDelta
 from openai.types.responses import Response
-from pydantic import BaseModel
+from pydantic import BaseModel, PrivateAttr, model_validator
 
 from crewai.events.types.llm_events import LLMCallType
-from crewai.llms.base_llm import BaseLLM, llm_call_context
+from crewai.llms.base_llm import BaseLLM, JsonResponseFormat, llm_call_context
+from crewai.llms.hooks.base import BaseInterceptor
 from crewai.llms.hooks.transport import AsyncHTTPTransport, HTTPTransport
 from crewai.utilities.agent_utils import is_context_length_exceeded
 from crewai.utilities.exceptions.context_window_exceeding_exception import (
@@ -29,7 +30,6 @@ from crewai.utilities.types import LLMMessage
 
 if TYPE_CHECKING:
     from crewai.agent.core import Agent
-    from crewai.llms.hooks.base import BaseInterceptor
     from crewai.task import Task
     from crewai.tools.base_tool import BaseTool
 
@@ -183,77 +183,69 @@ class OpenAICompletion(BaseLLM):
         "computer_use": "computer_use_preview",
     }
 
-    def __init__(
-        self,
-        model: str = "gpt-4o",
-        api_key: str | None = None,
-        base_url: str | None = None,
-        organization: str | None = None,
-        project: str | None = None,
-        timeout: float | None = None,
-        max_retries: int = 2,
-        default_headers: dict[str, str] | None = None,
-        default_query: dict[str, Any] | None = None,
-        client_params: dict[str, Any] | None = None,
-        temperature: float | None = None,
-        top_p: float | None = None,
-        frequency_penalty: float | None = None,
-        presence_penalty: float | None = None,
-        max_tokens: int | None = None,
-        max_completion_tokens: int | None = None,
-        seed: int | None = None,
-        stream: bool = False,
-        response_format: dict[str, Any] | type[BaseModel] | None = None,
-        logprobs: bool | None = None,
-        top_logprobs: int | None = None,
-        reasoning_effort: str | None = None,
-        provider: str | None = None,
-        interceptor: BaseInterceptor[httpx.Request, httpx.Response] | None = None,
-        api: Literal["completions", "responses"] = "completions",
-        instructions: str | None = None,
-        store: bool | None = None,
-        previous_response_id: str | None = None,
-        include: list[str] | None = None,
-        builtin_tools: list[str] | None = None,
-        parse_tool_outputs: bool = False,
-        auto_chain: bool = False,
-        auto_chain_reasoning: bool = False,
-        **kwargs: Any,
-    ) -> None:
-        """Initialize OpenAI completion client."""
+    model: str = "gpt-4o"
+    organization: str | None = None
+    project: str | None = None
+    timeout: float | None = None
+    max_retries: int = 2
+    default_headers: dict[str, str] | None = None
+    default_query: dict[str, Any] | None = None
+    client_params: dict[str, Any] | None = None
+    top_p: float | None = None
+    frequency_penalty: float | None = None
+    presence_penalty: float | None = None
+    max_tokens: int | None = None
+    max_completion_tokens: int | None = None
+    seed: int | None = None
+    stream: bool = False
+    response_format: JsonResponseFormat | type[BaseModel] | None = None
+    logprobs: bool | None = None
+    top_logprobs: int | None = None
+    reasoning_effort: str | None = None
+    interceptor: BaseInterceptor[httpx.Request, httpx.Response] | None = None
+    api: Literal["completions", "responses"] = "completions"
+    instructions: str | None = None
+    store: bool | None = None
+    previous_response_id: str | None = None
+    include: list[str] | None = None
+    builtin_tools: list[str] | None = None
+    parse_tool_outputs: bool = False
+    auto_chain: bool = False
+    auto_chain_reasoning: bool = False
+    api_base: str | None = None
+    is_o1_model: bool = False
+    is_gpt4_model: bool = False
 
-        if provider is None:
-            provider = kwargs.pop("provider", "openai")
+    _client: Any = PrivateAttr(default=None)
+    _async_client: Any = PrivateAttr(default=None)
+    _last_response_id: str | None = PrivateAttr(default=None)
+    _last_reasoning_items: list[Any] | None = PrivateAttr(default=None)
 
-        self.interceptor = interceptor
-        # Client configuration attributes
-        self.organization = organization
-        self.project = project
-        self.max_retries = max_retries
-        self.default_headers = default_headers
-        self.default_query = default_query
-        self.client_params = client_params
-        self.timeout = timeout
-        self.base_url = base_url
-        self.api_base = kwargs.pop("api_base", None)
-
-        super().__init__(
-            model=model,
-            temperature=temperature,
-            api_key=api_key or os.getenv("OPENAI_API_KEY"),
-            base_url=base_url,
-            timeout=timeout,
-            provider=provider,
-            **kwargs,
-        )
+    @model_validator(mode="before")
+    @classmethod
+    def _normalize_openai_fields(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
+        if not data.get("provider"):
+            data["provider"] = "openai"
+        data["api_key"] = data.get("api_key") or os.getenv("OPENAI_API_KEY")
+        # Extract api_base from kwargs if present
+        if "api_base" not in data:
+            data["api_base"] = None
+        model = data.get("model", "gpt-4o")
+        data["is_o1_model"] = "o1" in model.lower()
+        data["is_gpt4_model"] = "gpt-4" in model.lower()
+        return data
 
+    @model_validator(mode="after")
+    def _init_clients(self) -> OpenAICompletion:
         client_config = self._get_client_params()
         if self.interceptor:
             transport = HTTPTransport(interceptor=self.interceptor)
             http_client = httpx.Client(transport=transport)
             client_config["http_client"] = http_client
 
-        self.client = OpenAI(**client_config)
+        self._client = OpenAI(**client_config)
 
         async_client_config = self._get_client_params()
         if self.interceptor:
@@ -261,35 +253,8 @@ class OpenAICompletion(BaseLLM):
             async_http_client = httpx.AsyncClient(transport=async_transport)
             async_client_config["http_client"] = async_http_client
 
-        self.async_client = AsyncOpenAI(**async_client_config)
-
-        # Completion parameters
-        self.top_p = top_p
-        self.frequency_penalty = frequency_penalty
-        self.presence_penalty = presence_penalty
-        self.max_tokens = max_tokens
-        self.max_completion_tokens = max_completion_tokens
-        self.seed = seed
-        self.stream = stream
-        self.response_format = response_format
-        self.logprobs = logprobs
-        self.top_logprobs = top_logprobs
-        self.reasoning_effort = reasoning_effort
-        self.is_o1_model = "o1" in model.lower()
-        self.is_gpt4_model = "gpt-4" in model.lower()
-
-        # API selection and Responses API parameters
-        self.api = api
-        self.instructions = instructions
-        self.store = store
-        self.previous_response_id = previous_response_id
-        self.include = include
-        self.builtin_tools = builtin_tools
-        self.parse_tool_outputs = parse_tool_outputs
-        self.auto_chain = auto_chain
-        self.auto_chain_reasoning = auto_chain_reasoning
-        self._last_response_id: str | None = None
-        self._last_reasoning_items: list[Any] | None = None
+        self._async_client = AsyncOpenAI(**async_client_config)
+        return self
 
     @property
     def last_response_id(self) -> str | None:
@@ -818,7 +783,7 @@ class OpenAICompletion(BaseLLM):
     ) -> str | ResponsesAPIResult | Any:
         """Handle non-streaming Responses API call."""
         try:
-            response: Response = self.client.responses.create(**params)
+            response: Response = self._client.responses.create(**params)
 
             # Track response ID for auto-chaining
             if self.auto_chain and response.id:
@@ -844,6 +809,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("input", []),
+                    usage=usage,
                 )
 
                 return parsed_result
@@ -856,6 +822,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("input", []),
+                    usage=usage,
                 )
                 return function_calls
 
@@ -893,6 +860,7 @@ class OpenAICompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params.get("input", []),
+                        usage=usage,
                     )
                     return structured_result
                 except ValueError as e:
@@ -906,6 +874,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params.get("input", []),
+                usage=usage,
             )
 
             content = self._invoke_after_llm_call_hooks(
@@ -950,7 +919,7 @@ class OpenAICompletion(BaseLLM):
     ) -> str | ResponsesAPIResult | Any:
         """Handle async non-streaming Responses API call."""
         try:
-            response: Response = await self.async_client.responses.create(**params)
+            response: Response = await self._async_client.responses.create(**params)
 
             # Track response ID for auto-chaining
             if self.auto_chain and response.id:
@@ -976,6 +945,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("input", []),
+                    usage=usage,
                 )
 
                 return parsed_result
@@ -988,6 +958,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("input", []),
+                    usage=usage,
                 )
                 return function_calls
 
@@ -1025,6 +996,7 @@ class OpenAICompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params.get("input", []),
+                        usage=usage,
                     )
                     return structured_result
                 except ValueError as e:
@@ -1038,6 +1010,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params.get("input", []),
+                usage=usage,
             )
 
         except NotFoundError as e:
@@ -1080,8 +1053,9 @@ class OpenAICompletion(BaseLLM):
         full_response = ""
         function_calls: list[dict[str, Any]] = []
         final_response: Response | None = None
+        usage: dict[str, Any] | None = None
 
-        stream = self.client.responses.create(**params)
+        stream = self._client.responses.create(**params)
         response_id_stream = None
 
         for event in stream:
@@ -1137,6 +1111,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params.get("input", []),
+                usage=usage,
             )
 
             return parsed_result
@@ -1173,6 +1148,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("input", []),
+                    usage=usage,
                 )
                 return structured_result
             except ValueError as e:
@@ -1186,6 +1162,7 @@ class OpenAICompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params.get("input", []),
+            usage=usage,
         )
 
         return self._invoke_after_llm_call_hooks(
@@ -1204,8 +1181,9 @@ class OpenAICompletion(BaseLLM):
         full_response = ""
         function_calls: list[dict[str, Any]] = []
         final_response: Response | None = None
+        usage: dict[str, Any] | None = None
 
-        stream = await self.async_client.responses.create(**params)
+        stream = await self._async_client.responses.create(**params)
         response_id_stream = None
 
         async for event in stream:
@@ -1261,6 +1239,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params.get("input", []),
+                usage=usage,
             )
 
             return parsed_result
@@ -1297,6 +1276,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params.get("input", []),
+                    usage=usage,
                 )
                 return structured_result
             except ValueError as e:
@@ -1310,6 +1290,7 @@ class OpenAICompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params.get("input", []),
+            usage=usage,
         )
 
         return full_response
@@ -1595,7 +1576,7 @@ class OpenAICompletion(BaseLLM):
                 parse_params = {
                     k: v for k, v in params.items() if k != "response_format"
                 }
-                parsed_response = self.client.beta.chat.completions.parse(
+                parsed_response = self._client.beta.chat.completions.parse(
                     **parse_params,
                     response_format=response_model,
                 )
@@ -1615,10 +1596,11 @@ class OpenAICompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return parsed_object
 
-            response: ChatCompletion = self.client.chat.completions.create(**params)
+            response: ChatCompletion = self._client.chat.completions.create(**params)
 
             usage = self._extract_openai_token_usage(response)
 
@@ -1636,6 +1618,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage,
                 )
                 return list(message.tool_calls)
 
@@ -1674,6 +1657,7 @@ class OpenAICompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return structured_result
                 except ValueError as e:
@@ -1687,6 +1671,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage,
             )
 
             if usage.get("total_tokens", 0) > 0:
@@ -1728,7 +1713,7 @@ class OpenAICompletion(BaseLLM):
         self,
         full_response: str,
         tool_calls: dict[int, dict[str, Any]],
-        usage_data: dict[str, int],
+        usage_data: dict[str, Any] | None,
         params: dict[str, Any],
         available_functions: dict[str, Any] | None = None,
         from_task: Any | None = None,
@@ -1739,7 +1724,7 @@ class OpenAICompletion(BaseLLM):
         Args:
             full_response: The accumulated text response from the stream.
             tool_calls: Accumulated tool calls from the stream, keyed by index.
-            usage_data: Token usage data from the stream.
+            usage_data: Token usage data from the stream, or None if unavailable.
             params: The completion parameters containing messages.
             available_functions: Available functions for tool calling.
             from_task: Task that initiated the call.
@@ -1750,7 +1735,8 @@ class OpenAICompletion(BaseLLM):
             tool execution result when available_functions is provided,
             or the text response string.
         """
-        self._track_token_usage_internal(usage_data)
+        if usage_data:
+            self._track_token_usage_internal(usage_data)
 
         if tool_calls and not available_functions:
             tool_calls_list = [
@@ -1771,6 +1757,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage_data,
             )
             return tool_calls_list
 
@@ -1813,6 +1800,7 @@ class OpenAICompletion(BaseLLM):
             from_task=from_task,
             from_agent=from_agent,
             messages=params["messages"],
+            usage=usage_data,
         )
 
         return full_response
@@ -1837,7 +1825,7 @@ class OpenAICompletion(BaseLLM):
             }
 
             stream: ChatCompletionStream[BaseModel]
-            with self.client.beta.chat.completions.stream(
+            with self._client.beta.chat.completions.stream(
                 **parse_params, response_format=response_model
             ) as stream:
                 for chunk in stream:
@@ -1866,6 +1854,7 @@ class OpenAICompletion(BaseLLM):
                                 from_task=from_task,
                                 from_agent=from_agent,
                                 messages=params["messages"],
+                                usage=usage,
                             )
                             return parsed_result
 
@@ -1873,10 +1862,10 @@ class OpenAICompletion(BaseLLM):
             return ""
 
         completion_stream: Stream[ChatCompletionChunk] = (
-            self.client.chat.completions.create(**params)
+            self._client.chat.completions.create(**params)
         )
 
-        usage_data = {"total_tokens": 0}
+        usage_data: dict[str, Any] | None = None
 
         for completion_chunk in completion_stream:
             response_id_stream = (
@@ -1970,7 +1959,7 @@ class OpenAICompletion(BaseLLM):
                 parse_params = {
                     k: v for k, v in params.items() if k != "response_format"
                 }
-                parsed_response = await self.async_client.beta.chat.completions.parse(
+                parsed_response = await self._async_client.beta.chat.completions.parse(
                     **parse_params,
                     response_format=response_model,
                 )
@@ -1990,10 +1979,11 @@ class OpenAICompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return parsed_object
 
-            response: ChatCompletion = await self.async_client.chat.completions.create(
+            response: ChatCompletion = await self._async_client.chat.completions.create(
                 **params
             )
 
@@ -2013,6 +2003,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage,
                 )
                 return list(message.tool_calls)
 
@@ -2051,6 +2042,7 @@ class OpenAICompletion(BaseLLM):
                         from_task=from_task,
                         from_agent=from_agent,
                         messages=params["messages"],
+                        usage=usage,
                     )
                     return structured_result
                 except ValueError as e:
@@ -2064,6 +2056,7 @@ class OpenAICompletion(BaseLLM):
                 from_task=from_task,
                 from_agent=from_agent,
                 messages=params["messages"],
+                usage=usage,
             )
 
             if usage.get("total_tokens", 0) > 0:
@@ -2111,10 +2104,10 @@ class OpenAICompletion(BaseLLM):
         if response_model:
             completion_stream: AsyncIterator[
                 ChatCompletionChunk
-            ] = await self.async_client.chat.completions.create(**params)
+            ] = await self._async_client.chat.completions.create(**params)
 
             accumulated_content = ""
-            usage_data = {"total_tokens": 0}
+            usage_data: dict[str, Any] | None = None
             async for chunk in completion_stream:
                 response_id_stream = chunk.id if hasattr(chunk, "id") else None
 
@@ -2137,7 +2130,8 @@ class OpenAICompletion(BaseLLM):
                         response_id=response_id_stream,
                     )
 
-            self._track_token_usage_internal(usage_data)
+            if usage_data:
+                self._track_token_usage_internal(usage_data)
 
             try:
                 parsed_object = response_model.model_validate_json(accumulated_content)
@@ -2148,6 +2142,7 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage_data,
                 )
 
                 return parsed_object
@@ -2159,14 +2154,15 @@ class OpenAICompletion(BaseLLM):
                     from_task=from_task,
                     from_agent=from_agent,
                     messages=params["messages"],
+                    usage=usage_data,
                 )
                 return accumulated_content
 
         stream: AsyncIterator[
             ChatCompletionChunk
-        ] = await self.async_client.chat.completions.create(**params)
+        ] = await self._async_client.chat.completions.create(**params)
 
-        usage_data = {"total_tokens": 0}
+        usage_data = None
 
         async for chunk in stream:
             response_id_stream = chunk.id if hasattr(chunk, "id") else None
@@ -2245,6 +2241,9 @@ class OpenAICompletion(BaseLLM):
 
     def supports_stop_words(self) -> bool:
         """Check if the model supports stop words."""
+        model_lower = self.model.lower() if self.model else ""
+        if "gpt-5" in model_lower:
+            return False
         return not self.is_o1_model
 
     def get_context_window_size(self) -> int:
@@ -2353,8 +2352,8 @@ class OpenAICompletion(BaseLLM):
             from crewai_files.uploaders.openai import OpenAIFileUploader
 
             return OpenAIFileUploader(
-                client=self.client,
-                async_client=self.async_client,
+                client=self._client,
+                async_client=self._async_client,
             )
         except ImportError:
             return None

lib/crewai/src/crewai/llms/providers/openai_compatible/completion.py

@@ -16,6 +16,8 @@ from dataclasses import dataclass, field
 import os
 from typing import Any
 
+from pydantic import model_validator
+
 from crewai.llms.providers.openai.completion import OpenAICompletion
 
 
@@ -140,31 +142,13 @@ class OpenAICompatibleCompletion(OpenAICompletion):
         )
     """
 
-    def __init__(
-        self,
-        model: str,
-        provider: str,
-        api_key: str | None = None,
-        base_url: str | None = None,
-        default_headers: dict[str, str] | None = None,
-        **kwargs: Any,
-    ) -> None:
-        """Initialize OpenAI-compatible completion client.
+    @model_validator(mode="before")
+    @classmethod
+    def _resolve_provider_config(cls, data: Any) -> Any:
+        if not isinstance(data, dict):
+            return data
 
-        Args:
-            model: The model identifier.
-            provider: The provider name (must be in OPENAI_COMPATIBLE_PROVIDERS).
-            api_key: Optional API key override. If not provided, uses the
-                provider's configured environment variable.
-            base_url: Optional base URL override. If not provided, uses the
-                provider's configured default or environment variable.
-            default_headers: Optional headers to merge with provider defaults.
-            **kwargs: Additional arguments passed to OpenAICompletion.
-
-        Raises:
-            ValueError: If the provider is not supported or required API key
-                is missing.
-        """
+        provider = data.get("provider", "")
         config = OPENAI_COMPATIBLE_PROVIDERS.get(provider)
         if config is None:
             supported = ", ".join(sorted(OPENAI_COMPATIBLE_PROVIDERS.keys()))
@@ -173,21 +157,15 @@ class OpenAICompatibleCompletion(OpenAICompletion):
                 f"Supported providers: {supported}"
             )
 
-        resolved_api_key = self._resolve_api_key(api_key, config, provider)
-        resolved_base_url = self._resolve_base_url(base_url, config, provider)
-        resolved_headers = self._resolve_headers(default_headers, config)
-
-        super().__init__(
-            model=model,
-            provider=provider,
-            api_key=resolved_api_key,
-            base_url=resolved_base_url,
-            default_headers=resolved_headers,
-            **kwargs,
+        data["api_key"] = cls._resolve_api_key(data.get("api_key"), config, provider)
+        data["base_url"] = cls._resolve_base_url(data.get("base_url"), config, provider)
+        data["default_headers"] = cls._resolve_headers(
+            data.get("default_headers"), config
         )
+        return data
 
+    @staticmethod
     def _resolve_api_key(
-        self,
         api_key: str | None,
         config: ProviderConfig,
         provider: str,
@@ -220,8 +198,8 @@ class OpenAICompatibleCompletion(OpenAICompletion):
 
         return config.default_api_key
 
+    @staticmethod
     def _resolve_base_url(
-        self,
         base_url: str | None,
         config: ProviderConfig,
         provider: str,
@@ -249,8 +227,8 @@ class OpenAICompatibleCompletion(OpenAICompletion):
 
         return resolved
 
+    @staticmethod
     def _resolve_headers(
-        self,
         headers: dict[str, str] | None,
         config: ProviderConfig,
     ) -> dict[str, str] | None:

lib/crewai/src/crewai/llms/third_party/__init__.py

@@ -1 +0,0 @@
-"""Third-party LLM implementations for crewAI."""

lib/crewai/src/crewai/memory/encoding_flow.py

@@ -98,7 +98,7 @@ class EncodingFlow(Flow[EncodingState]):
 
     _skip_auto_memory: bool = True
 
-    initial_state = EncodingState
+    initial_state: type[EncodingState] = EncodingState
 
     def __init__(
         self,

lib/crewai/src/crewai/memory/recall_flow.py

@@ -65,7 +65,7 @@ class RecallFlow(Flow[RecallState]):
 
     _skip_auto_memory: bool = True
 
-    initial_state = RecallState
+    initial_state: type[RecallState] = RecallState
 
     def __init__(
         self,

lib/crewai/src/crewai/memory/unified_memory.py

@@ -148,6 +148,36 @@ class Memory(BaseModel):
     _pending_saves: list[Future[Any]] = PrivateAttr(default_factory=list)
     _pending_lock: threading.Lock = PrivateAttr(default_factory=threading.Lock)
 
+    def __deepcopy__(self, memo: dict[int, Any] | None = None) -> Memory:
+        """Deepcopy that handles unpickleable private attrs (ThreadPoolExecutor, Lock)."""
+        import copy as _copy
+
+        cls = type(self)
+        new = cls.__new__(cls)
+        if memo is None:
+            memo = {}
+        memo[id(self)] = new
+        object.__setattr__(new, "__dict__", _copy.deepcopy(self.__dict__, memo))
+        object.__setattr__(
+            new, "__pydantic_fields_set__", _copy.copy(self.__pydantic_fields_set__)
+        )
+        object.__setattr__(
+            new, "__pydantic_extra__", _copy.deepcopy(self.__pydantic_extra__, memo)
+        )
+        # Private attrs: create fresh pool/lock instead of deepcopying
+        private = {}
+        for k, v in (self.__pydantic_private__ or {}).items():
+            if isinstance(v, (ThreadPoolExecutor, threading.Lock)):
+                attr = self.__private_attributes__[k]
+                private[k] = attr.get_default()
+            else:
+                try:
+                    private[k] = _copy.deepcopy(v, memo)
+                except Exception:
+                    private[k] = v
+        object.__setattr__(new, "__pydantic_private__", private)
+        return new
+
     def model_post_init(self, __context: Any) -> None:
         """Initialize runtime state from field values."""
         self._config = MemoryConfig(

lib/crewai/src/crewai/utilities/evaluators/crew_evaluator_handler.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 from collections import defaultdict
 from typing import TYPE_CHECKING, Any
 
-from pydantic import BaseModel, Field, InstanceOf
+from pydantic import BaseModel, Field
 from rich.box import HEAVY_EDGE
 from rich.console import Console
 from rich.table import Table
@@ -39,9 +39,9 @@ class CrewEvaluator:
     def __init__(
         self,
         crew: Crew,
-        eval_llm: InstanceOf[BaseLLM] | str | None = None,
+        eval_llm: BaseLLM | str | None = None,
         openai_model_name: str | None = None,
-        llm: InstanceOf[BaseLLM] | str | None = None,
+        llm: BaseLLM | str | None = None,
     ) -> None:
         self.crew = crew
         self.llm = eval_llm

lib/crewai/src/crewai/utilities/prompts.py

@@ -2,9 +2,10 @@
 
 from __future__ import annotations
 
-from typing import Annotated, Any, Literal, TypedDict
+from typing import Annotated, Any, Literal
 
 from pydantic import BaseModel, Field
+from typing_extensions import TypedDict
 
 from crewai.utilities.i18n import I18N, get_i18n
 

lib/crewai/tests/agents/test_agent.py

@@ -1692,9 +1692,27 @@ def test_agent_with_knowledge_sources_works_with_copy():
         ) as mock_knowledge_storage:
             from crewai.knowledge.storage.base_knowledge_storage import BaseKnowledgeStorage
 
-            mock_knowledge_storage_instance = mock_knowledge_storage.return_value
-            mock_knowledge_storage_instance.__class__ = BaseKnowledgeStorage
-            agent.knowledge_storage = mock_knowledge_storage_instance
+            class _StubStorage(BaseKnowledgeStorage):
+                def search(self, query, limit=5, metadata_filter=None, score_threshold=0.6):
+                    return []
+
+                async def asearch(self, query, limit=5, metadata_filter=None, score_threshold=0.6):
+                    return []
+
+                def save(self, documents):
+                    pass
+
+                async def asave(self, documents):
+                    pass
+
+                def reset(self):
+                    pass
+
+                async def areset(self):
+                    pass
+
+            mock_knowledge_storage.return_value = _StubStorage()
+            agent.knowledge_storage = _StubStorage()
 
             agent_copy = agent.copy()
 

lib/crewai/tests/agents/test_agent_executor.py

@@ -4,13 +4,55 @@ Tests the Flow-based agent executor implementation including state management,
 flow methods, routing logic, and error handling.
 """
 
+from __future__ import annotations
+
 import asyncio
 import time
+from typing import Any
 from unittest.mock import AsyncMock, Mock, patch
 
 import pytest
 
+from crewai.agents.tools_handler import ToolsHandler as _ToolsHandler
 from crewai.agents.step_executor import StepExecutor
+
+
+def _build_executor(**kwargs: Any) -> AgentExecutor:
+    """Create an AgentExecutor without validation — for unit tests.
+
+    Uses model_construct to skip Pydantic validators so plain Mock()
+    objects are accepted for typed fields like llm, agent, crew, task.
+    """
+    executor = AgentExecutor.model_construct(**kwargs)
+    executor._state = AgentExecutorState()
+    executor._methods = {}
+    executor._method_outputs = []
+    executor._completed_methods = set()
+    executor._fired_or_listeners = set()
+    executor._pending_and_listeners = {}
+    executor._method_execution_counts = {}
+    executor._method_call_counts = {}
+    executor._event_futures = []
+    executor._human_feedback_method_outputs = {}
+    executor._input_history = []
+    executor._is_execution_resuming = False
+    import threading
+    executor._state_lock = threading.Lock()
+    executor._or_listeners_lock = threading.Lock()
+    executor._execution_lock = threading.Lock()
+    executor._finalize_lock = threading.Lock()
+    executor._finalize_called = False
+    executor._is_executing = False
+    executor._has_been_invoked = False
+    executor._last_parser_error = None
+    executor._last_context_error = None
+    executor._step_executor = None
+    executor._planner_observer = None
+    from crewai.utilities.printer import Printer
+    executor._printer = Printer()
+    from crewai.utilities.i18n import get_i18n
+    executor._i18n = kwargs.get("i18n") or get_i18n()
+    return executor
 from crewai.agents.planner_observer import PlannerObserver
 from crewai.experimental.agent_executor import (
     AgentExecutorState,
@@ -75,6 +117,7 @@ class TestAgentExecutor:
         """Create mock dependencies for executor."""
         llm = Mock()
         llm.supports_stop_words.return_value = True
+        llm.stop = []
 
         task = Mock()
         task.description = "Test task"
@@ -94,7 +137,7 @@ class TestAgentExecutor:
         prompt = {"prompt": "Test prompt with {input}, {tool_names}, {tools}"}
 
         tools = []
-        tools_handler = Mock()
+        tools_handler = Mock(spec=_ToolsHandler)
 
         return {
             "llm": llm,
@@ -112,7 +155,7 @@ class TestAgentExecutor:
 
     def test_executor_initialization(self, mock_dependencies):
         """Test AgentExecutor initialization."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         assert executor.llm == mock_dependencies["llm"]
         assert executor.task == mock_dependencies["task"]
@@ -126,7 +169,7 @@ class TestAgentExecutor:
         with patch.object(
             AgentExecutor, "_show_start_logs"
         ) as mock_show_start:
-            executor = AgentExecutor(**mock_dependencies)
+            executor = _build_executor(**mock_dependencies)
             result = executor.initialize_reasoning()
 
             assert result == "initialized"
@@ -134,7 +177,7 @@ class TestAgentExecutor:
 
     def test_check_max_iterations_not_reached(self, mock_dependencies):
         """Test routing when iterations < max."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.iterations = 5
 
         result = executor.check_max_iterations()
@@ -142,7 +185,7 @@ class TestAgentExecutor:
 
     def test_check_max_iterations_reached(self, mock_dependencies):
         """Test routing when iterations >= max."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.iterations = 10
 
         result = executor.check_max_iterations()
@@ -150,7 +193,7 @@ class TestAgentExecutor:
 
     def test_route_by_answer_type_action(self, mock_dependencies):
         """Test routing for AgentAction."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.current_answer = AgentAction(
             thought="thinking", tool="search", tool_input="query", text="action text"
         )
@@ -160,7 +203,7 @@ class TestAgentExecutor:
 
     def test_route_by_answer_type_finish(self, mock_dependencies):
         """Test routing for AgentFinish."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.current_answer = AgentFinish(
             thought="final thoughts", output="Final answer", text="complete"
         )
@@ -170,7 +213,7 @@ class TestAgentExecutor:
 
     def test_continue_iteration(self, mock_dependencies):
         """Test iteration continuation."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         result = executor.continue_iteration()
 
@@ -179,7 +222,7 @@ class TestAgentExecutor:
     def test_finalize_success(self, mock_dependencies):
         """Test finalize with valid AgentFinish."""
         with patch.object(AgentExecutor, "_show_logs") as mock_show_logs:
-            executor = AgentExecutor(**mock_dependencies)
+            executor = _build_executor(**mock_dependencies)
             executor.state.current_answer = AgentFinish(
                 thought="final thinking", output="Done", text="complete"
             )
@@ -192,7 +235,7 @@ class TestAgentExecutor:
 
     def test_finalize_failure(self, mock_dependencies):
         """Test finalize skips when given AgentAction instead of AgentFinish."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.current_answer = AgentAction(
             thought="thinking", tool="search", tool_input="query", text="action text"
         )
@@ -208,7 +251,7 @@ class TestAgentExecutor:
     ):
         """Finalize should skip synthesis when last todo is already a complete answer."""
         with patch.object(AgentExecutor, "_show_logs") as mock_show_logs:
-            executor = AgentExecutor(**mock_dependencies)
+            executor = _build_executor(**mock_dependencies)
             executor.state.todos.items = [
                 TodoItem(
                     step_number=1,
@@ -252,7 +295,7 @@ class TestAgentExecutor:
     ):
         """Finalize should still synthesize when response_model is configured."""
         with patch.object(AgentExecutor, "_show_logs"):
-            executor = AgentExecutor(**mock_dependencies)
+            executor = _build_executor(**mock_dependencies)
             executor.response_model = Mock()
             executor.state.todos.items = [
                 TodoItem(
@@ -287,7 +330,7 @@ class TestAgentExecutor:
 
     def test_format_prompt(self, mock_dependencies):
         """Test prompt formatting."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         inputs = {"input": "test input", "tool_names": "tool1, tool2", "tools": "desc"}
 
         result = executor._format_prompt("Prompt {input} {tool_names} {tools}", inputs)
@@ -298,18 +341,18 @@ class TestAgentExecutor:
 
     def test_is_training_mode_false(self, mock_dependencies):
         """Test training mode detection when not in training."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         assert executor._is_training_mode() is False
 
     def test_is_training_mode_true(self, mock_dependencies):
         """Test training mode detection when in training."""
         mock_dependencies["crew"]._train = True
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         assert executor._is_training_mode() is True
 
     def test_append_message_to_state(self, mock_dependencies):
         """Test message appending to state."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         initial_count = len(executor.state.messages)
 
         executor._append_message_to_state("test message")
@@ -322,7 +365,7 @@ class TestAgentExecutor:
         callback = Mock()
         mock_dependencies["step_callback"] = callback
 
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         answer = AgentFinish(thought="thinking", output="test", text="final")
 
         executor._invoke_step_callback(answer)
@@ -332,7 +375,7 @@ class TestAgentExecutor:
     def test_invoke_step_callback_none(self, mock_dependencies):
         """Test step callback when none provided."""
         mock_dependencies["step_callback"] = None
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         # Should not raise error
         executor._invoke_step_callback(
@@ -346,7 +389,7 @@ class TestAgentExecutor:
         """Test async step callback scheduling when already in an event loop."""
         callback = AsyncMock()
         mock_dependencies["step_callback"] = callback
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         answer = AgentFinish(thought="thinking", output="test", text="final")
         with patch("crewai.experimental.agent_executor.asyncio.run") as mock_run:
@@ -364,6 +407,7 @@ class TestStepExecutorCriticalFixes:
     def mock_dependencies(self):
         """Create mock dependencies for AgentExecutor tests in this class."""
         llm = Mock()
+        llm.stop = []
         llm.supports_stop_words.return_value = True
 
         task = Mock()
@@ -393,6 +437,7 @@ class TestStepExecutorCriticalFixes:
     @pytest.fixture
     def step_executor(self):
         llm = Mock()
+        llm.stop = []
         llm.supports_stop_words.return_value = True
 
         agent = Mock()
@@ -485,7 +530,7 @@ class TestStepExecutorCriticalFixes:
 
         mock_handle_exception.return_value = None
 
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor._last_parser_error = OutputParserError("test error")
         initial_iterations = executor.state.iterations
 
@@ -500,7 +545,7 @@ class TestStepExecutorCriticalFixes:
         self, mock_handle_context, mock_dependencies
     ):
         """Test recovery from context length error."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor._last_context_error = Exception("context too long")
         initial_iterations = executor.state.iterations
 
@@ -513,16 +558,16 @@ class TestStepExecutorCriticalFixes:
     def test_use_stop_words_property(self, mock_dependencies):
         """Test use_stop_words property."""
         mock_dependencies["llm"].supports_stop_words.return_value = True
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         assert executor.use_stop_words is True
 
         mock_dependencies["llm"].supports_stop_words.return_value = False
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         assert executor.use_stop_words is False
 
     def test_compatibility_properties(self, mock_dependencies):
         """Test compatibility properties for mixin."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.messages = [{"role": "user", "content": "test"}]
         executor.state.iterations = 5
 
@@ -538,6 +583,7 @@ class TestFlowErrorHandling:
     def mock_dependencies(self):
         """Create mock dependencies."""
         llm = Mock()
+        llm.stop = []
         llm.supports_stop_words.return_value = True
 
         task = Mock()
@@ -575,7 +621,7 @@ class TestFlowErrorHandling:
         mock_enforce_rpm.return_value = None
         mock_get_llm.side_effect = OutputParserError("parse failed")
 
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         result = executor.call_llm_and_parse()
 
         assert result == "parser_error"
@@ -596,7 +642,7 @@ class TestFlowErrorHandling:
         mock_get_llm.side_effect = Exception("context length")
         mock_is_context_exceeded.return_value = True
 
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         result = executor.call_llm_and_parse()
 
         assert result == "context_error"
@@ -610,6 +656,7 @@ class TestFlowInvoke:
     def mock_dependencies(self):
         """Create mock dependencies."""
         llm = Mock()
+        llm.stop = []
         task = Mock()
         task.description = "Test"
         task.human_input = False
@@ -646,7 +693,7 @@ class TestFlowInvoke:
         mock_dependencies,
     ):
         """Test successful invoke without human feedback."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         # Mock kickoff to set the final answer in state
         def mock_kickoff_side_effect():
@@ -666,7 +713,7 @@ class TestFlowInvoke:
     @patch.object(AgentExecutor, "kickoff")
     def test_invoke_failure_no_agent_finish(self, mock_kickoff, mock_dependencies):
         """Test invoke fails without AgentFinish."""
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         executor.state.current_answer = AgentAction(
             thought="thinking", tool="test", tool_input="test", text="action text"
         )
@@ -689,7 +736,7 @@ class TestFlowInvoke:
             "system": "System: {input}",
             "user": "User: {input} {tool_names} {tools}",
         }
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         def mock_kickoff_side_effect():
             executor.state.current_answer = AgentFinish(
@@ -713,6 +760,7 @@ class TestNativeToolExecution:
     @pytest.fixture
     def mock_dependencies(self):
         llm = Mock()
+        llm.stop = []
         llm.supports_stop_words.return_value = True
 
         task = Mock()
@@ -734,7 +782,7 @@ class TestNativeToolExecution:
 
         prompt = {"prompt": "Test {input} {tool_names} {tools}"}
 
-        tools_handler = Mock()
+        tools_handler = Mock(spec=_ToolsHandler)
         tools_handler.cache = None
 
         return {
@@ -754,7 +802,7 @@ class TestNativeToolExecution:
     def test_execute_native_tool_runs_parallel_for_multiple_calls(
         self, mock_dependencies
     ):
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         def slow_one() -> str:
             time.sleep(0.2)
@@ -790,7 +838,7 @@ class TestNativeToolExecution:
     def test_execute_native_tool_falls_back_to_sequential_for_result_as_answer(
         self, mock_dependencies
     ):
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
 
         def slow_one() -> str:
             time.sleep(0.2)
@@ -832,7 +880,7 @@ class TestNativeToolExecution:
     def test_execute_native_tool_result_as_answer_short_circuits_remaining_calls(
         self, mock_dependencies
     ):
-        executor = AgentExecutor(**mock_dependencies)
+        executor = _build_executor(**mock_dependencies)
         call_counts = {"slow_one": 0, "slow_two": 0}
 
         def slow_one() -> str:
@@ -879,30 +927,6 @@ class TestNativeToolExecution:
         assert len(tool_messages) == 1
         assert tool_messages[0]["tool_call_id"] == "call_1"
 
-    def test_check_native_todo_completion_requires_current_todo(
-        self, mock_dependencies
-    ):
-        from crewai.utilities.planning_types import TodoList
-
-        executor = AgentExecutor(**mock_dependencies)
-
-        # No current todo → not satisfied
-        executor.state.todos = TodoList(items=[])
-        assert executor.check_native_todo_completion() == "todo_not_satisfied"
-
-        # With a current todo that has tool_to_use → satisfied
-        running = TodoItem(
-            step_number=1,
-            description="Use the expected tool",
-            tool_to_use="expected_tool",
-            status="running",
-        )
-        executor.state.todos = TodoList(items=[running])
-        assert executor.check_native_todo_completion() == "todo_satisfied"
-
-        # With a current todo without tool_to_use → still satisfied
-        running.tool_to_use = None
-        assert executor.check_native_todo_completion() == "todo_satisfied"
 
 
 class TestPlannerObserver:

lib/crewai/tests/cassettes/llms/openai/test_openai_responses_api_cached_prompt_tokens_with_tools.yaml

@@ -1,7 +1,11 @@
 interactions:
 - request:
-    body: '{"messages":[{"role":"system","content":"You are a helpful assistant that
-      uses tools. This is padding text to ensure the prompt is large enough for caching.
+    body: '{"input":[{"role":"user","content":"What is the weather in Tokyo?"}],"model":"gpt-4.1","instructions":"You
+      are a helpful assistant that uses tools. This is padding text to ensure the
+      prompt is large enough for caching. This is padding text to ensure the prompt
+      is large enough for caching. This is padding text to ensure the prompt is large
+      enough for caching. This is padding text to ensure the prompt is large enough
+      for caching. This is padding text to ensure the prompt is large enough for caching.
       This is padding text to ensure the prompt is large enough for caching. This
       is padding text to ensure the prompt is large enough for caching. This is padding
       text to ensure the prompt is large enough for caching. This is padding text
@@ -68,13 +72,9 @@ interactions:
       for caching. This is padding text to ensure the prompt is large enough for caching.
       This is padding text to ensure the prompt is large enough for caching. This
       is padding text to ensure the prompt is large enough for caching. This is padding
-      text to ensure the prompt is large enough for caching. This is padding text
-      to ensure the prompt is large enough for caching. This is padding text to ensure
-      the prompt is large enough for caching. This is padding text to ensure the prompt
-      is large enough for caching. This is padding text to ensure the prompt is large
-      enough for caching. "},{"role":"user","content":"What is the weather in Tokyo?"}],"model":"gpt-4.1","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_weather","description":"Get
-      the current weather for a location","strict":true,"parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-      city name"}},"required":["location"],"additionalProperties":false}}}]}'
+      text to ensure the prompt is large enough for caching. ","tools":[{"type":"function","name":"get_weather","description":"Get
+      the current weather for a location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
+      city name"}},"required":["location"]}}]}'
     headers:
       User-Agent:
       - X-USER-AGENT-XXX
@@ -87,7 +87,7 @@ interactions:
       connection:
       - keep-alive
       content-length:
-      - '6158'
+      - '6065'
       content-type:
       - application/json
       host:
@@ -109,26 +109,113 @@ interactions:
       x-stainless-runtime:
       - CPython
       x-stainless-runtime-version:
-      - 3.13.3
+      - 3.13.12
     method: POST
-    uri: https://api.openai.com/v1/chat/completions
+    uri: https://api.openai.com/v1/responses
   response:
     body:
-      string: "{\n  \"id\": \"chatcmpl-D7mXQCgT3p3ViImkiqDiZGqLREQtp\",\n  \"object\":
-        \"chat.completion\",\n  \"created\": 1770747248,\n  \"model\": \"gpt-4.1-2025-04-14\",\n
-        \ \"choices\": [\n    {\n      \"index\": 0,\n      \"message\": {\n        \"role\":
-        \"assistant\",\n        \"content\": null,\n        \"tool_calls\": [\n          {\n
-        \           \"id\": \"call_9ZqMavn3J1fBnQEaqpYol0Bd\",\n            \"type\":
-        \"function\",\n            \"function\": {\n              \"name\": \"get_weather\",\n
-        \             \"arguments\": \"{\\\"location\\\":\\\"Tokyo\\\"}\"\n            }\n
-        \         }\n        ],\n        \"refusal\": null,\n        \"annotations\":
-        []\n      },\n      \"logprobs\": null,\n      \"finish_reason\": \"tool_calls\"\n
-        \   }\n  ],\n  \"usage\": {\n    \"prompt_tokens\": 1187,\n    \"completion_tokens\":
-        14,\n    \"total_tokens\": 1201,\n    \"prompt_tokens_details\": {\n      \"cached_tokens\":
-        1152,\n      \"audio_tokens\": 0\n    },\n    \"completion_tokens_details\":
-        {\n      \"reasoning_tokens\": 0,\n      \"audio_tokens\": 0,\n      \"accepted_prediction_tokens\":
-        0,\n      \"rejected_prediction_tokens\": 0\n    }\n  },\n  \"service_tier\":
-        \"default\",\n  \"system_fingerprint\": \"fp_8b22347a3e\"\n}\n"
+      string: "{\n  \"id\": \"resp_0d68149bcc0d14810069caf464a4b48197bd9f098abb2f6303\",\n
+        \ \"object\": \"response\",\n  \"created_at\": 1774908516,\n  \"status\":
+        \"completed\",\n  \"background\": false,\n  \"billing\": {\n    \"payer\":
+        \"developer\"\n  },\n  \"completed_at\": 1774908517,\n  \"error\": null,\n
+        \ \"frequency_penalty\": 0.0,\n  \"incomplete_details\": null,\n  \"instructions\":
+        \"You are a helpful assistant that uses tools. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. \",\n  \"max_output_tokens\":
+        null,\n  \"max_tool_calls\": null,\n  \"model\": \"gpt-4.1-2025-04-14\",\n
+        \ \"output\": [\n    {\n      \"id\": \"fc_0d68149bcc0d14810069caf46568088197a33be67f16a1fa09\",\n
+        \     \"type\": \"function_call\",\n      \"status\": \"completed\",\n      \"arguments\":
+        \"{\\\"location\\\":\\\"Tokyo\\\"}\",\n      \"call_id\": \"call_74rwmYse0DE4JFaFGyAFx9bu\",\n
+        \     \"name\": \"get_weather\"\n    }\n  ],\n  \"parallel_tool_calls\": true,\n
+        \ \"presence_penalty\": 0.0,\n  \"previous_response_id\": null,\n  \"prompt_cache_key\":
+        null,\n  \"prompt_cache_retention\": null,\n  \"reasoning\": {\n    \"effort\":
+        null,\n    \"summary\": null\n  },\n  \"safety_identifier\": null,\n  \"service_tier\":
+        \"default\",\n  \"store\": true,\n  \"temperature\": 1.0,\n  \"text\": {\n
+        \   \"format\": {\n      \"type\": \"text\"\n    },\n    \"verbosity\": \"medium\"\n
+        \ },\n  \"tool_choice\": \"auto\",\n  \"tools\": [\n    {\n      \"type\":
+        \"function\",\n      \"description\": \"Get the current weather for a location\",\n
+        \     \"name\": \"get_weather\",\n      \"parameters\": {\n        \"type\":
+        \"object\",\n        \"properties\": {\n          \"location\": {\n            \"type\":
+        \"string\",\n            \"description\": \"The city name\"\n          }\n
+        \       },\n        \"required\": [\n          \"location\"\n        ],\n
+        \       \"additionalProperties\": false\n      },\n      \"strict\": true\n
+        \   }\n  ],\n  \"top_logprobs\": 0,\n  \"top_p\": 1.0,\n  \"truncation\":
+        \"disabled\",\n  \"usage\": {\n    \"input_tokens\": 1185,\n    \"input_tokens_details\":
+        {\n      \"cached_tokens\": 0\n    },\n    \"output_tokens\": 15,\n    \"output_tokens_details\":
+        {\n      \"reasoning_tokens\": 0\n    },\n    \"total_tokens\": 1200\n  },\n
+        \ \"user\": null,\n  \"metadata\": {}\n}"
     headers:
       CF-RAY:
       - CF-RAY-XXX
@@ -137,7 +224,7 @@ interactions:
       Content-Type:
       - application/json
       Date:
-      - Tue, 10 Feb 2026 18:14:08 GMT
+      - Mon, 30 Mar 2026 22:08:37 GMT
       Server:
       - cloudflare
       Strict-Transport-Security:
@@ -146,8 +233,6 @@ interactions:
       - chunked
       X-Content-Type-Options:
       - X-CONTENT-TYPE-XXX
-      access-control-expose-headers:
-      - ACCESS-CONTROL-XXX
       alt-svc:
       - h3=":443"; ma=86400
       cf-cache-status:
@@ -155,15 +240,13 @@ interactions:
       openai-organization:
       - OPENAI-ORG-XXX
       openai-processing-ms:
-      - '484'
+      - '1085'
       openai-project:
       - OPENAI-PROJECT-XXX
       openai-version:
       - '2020-10-01'
       set-cookie:
       - SET-COOKIE-XXX
-      x-openai-proxy-wasm:
-      - v0.1
       x-ratelimit-limit-requests:
       - X-RATELIMIT-LIMIT-REQUESTS-XXX
       x-ratelimit-limit-tokens:
@@ -182,8 +265,12 @@ interactions:
       code: 200
       message: OK
 - request:
-    body: '{"messages":[{"role":"system","content":"You are a helpful assistant that
-      uses tools. This is padding text to ensure the prompt is large enough for caching.
+    body: '{"input":[{"role":"user","content":"What is the weather in Paris?"}],"model":"gpt-4.1","instructions":"You
+      are a helpful assistant that uses tools. This is padding text to ensure the
+      prompt is large enough for caching. This is padding text to ensure the prompt
+      is large enough for caching. This is padding text to ensure the prompt is large
+      enough for caching. This is padding text to ensure the prompt is large enough
+      for caching. This is padding text to ensure the prompt is large enough for caching.
       This is padding text to ensure the prompt is large enough for caching. This
       is padding text to ensure the prompt is large enough for caching. This is padding
       text to ensure the prompt is large enough for caching. This is padding text
@@ -250,13 +337,9 @@ interactions:
       for caching. This is padding text to ensure the prompt is large enough for caching.
       This is padding text to ensure the prompt is large enough for caching. This
       is padding text to ensure the prompt is large enough for caching. This is padding
-      text to ensure the prompt is large enough for caching. This is padding text
-      to ensure the prompt is large enough for caching. This is padding text to ensure
-      the prompt is large enough for caching. This is padding text to ensure the prompt
-      is large enough for caching. This is padding text to ensure the prompt is large
-      enough for caching. "},{"role":"user","content":"What is the weather in Paris?"}],"model":"gpt-4.1","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_weather","description":"Get
-      the current weather for a location","strict":true,"parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-      city name"}},"required":["location"],"additionalProperties":false}}}]}'
+      text to ensure the prompt is large enough for caching. ","tools":[{"type":"function","name":"get_weather","description":"Get
+      the current weather for a location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
+      city name"}},"required":["location"]}}]}'
     headers:
       User-Agent:
       - X-USER-AGENT-XXX
@@ -269,7 +352,7 @@ interactions:
       connection:
       - keep-alive
       content-length:
-      - '6158'
+      - '6065'
       content-type:
       - application/json
       cookie:
@@ -293,26 +376,113 @@ interactions:
       x-stainless-runtime:
       - CPython
       x-stainless-runtime-version:
-      - 3.13.3
+      - 3.13.12
     method: POST
-    uri: https://api.openai.com/v1/chat/completions
+    uri: https://api.openai.com/v1/responses
   response:
     body:
-      string: "{\n  \"id\": \"chatcmpl-D7mXR8k9vk8TlGvGXlrQSI7iNeAN1\",\n  \"object\":
-        \"chat.completion\",\n  \"created\": 1770747249,\n  \"model\": \"gpt-4.1-2025-04-14\",\n
-        \ \"choices\": [\n    {\n      \"index\": 0,\n      \"message\": {\n        \"role\":
-        \"assistant\",\n        \"content\": null,\n        \"tool_calls\": [\n          {\n
-        \           \"id\": \"call_6PeUBlRPG8JcV2lspmLjJbnn\",\n            \"type\":
-        \"function\",\n            \"function\": {\n              \"name\": \"get_weather\",\n
-        \             \"arguments\": \"{\\\"location\\\":\\\"Paris\\\"}\"\n            }\n
-        \         }\n        ],\n        \"refusal\": null,\n        \"annotations\":
-        []\n      },\n      \"logprobs\": null,\n      \"finish_reason\": \"tool_calls\"\n
-        \   }\n  ],\n  \"usage\": {\n    \"prompt_tokens\": 1187,\n    \"completion_tokens\":
-        14,\n    \"total_tokens\": 1201,\n    \"prompt_tokens_details\": {\n      \"cached_tokens\":
-        1152,\n      \"audio_tokens\": 0\n    },\n    \"completion_tokens_details\":
-        {\n      \"reasoning_tokens\": 0,\n      \"audio_tokens\": 0,\n      \"accepted_prediction_tokens\":
-        0,\n      \"rejected_prediction_tokens\": 0\n    }\n  },\n  \"service_tier\":
-        \"default\",\n  \"system_fingerprint\": \"fp_8b22347a3e\"\n}\n"
+      string: "{\n  \"id\": \"resp_0525bf798202137e0069caf465ee3c8196aa7c83da1c369eb7\",\n
+        \ \"object\": \"response\",\n  \"created_at\": 1774908517,\n  \"status\":
+        \"completed\",\n  \"background\": false,\n  \"billing\": {\n    \"payer\":
+        \"developer\"\n  },\n  \"completed_at\": 1774908518,\n  \"error\": null,\n
+        \ \"frequency_penalty\": 0.0,\n  \"incomplete_details\": null,\n  \"instructions\":
+        \"You are a helpful assistant that uses tools. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. This is
+        padding text to ensure the prompt is large enough for caching. This is padding
+        text to ensure the prompt is large enough for caching. This is padding text
+        to ensure the prompt is large enough for caching. This is padding text to
+        ensure the prompt is large enough for caching. This is padding text to ensure
+        the prompt is large enough for caching. This is padding text to ensure the
+        prompt is large enough for caching. This is padding text to ensure the prompt
+        is large enough for caching. This is padding text to ensure the prompt is
+        large enough for caching. This is padding text to ensure the prompt is large
+        enough for caching. This is padding text to ensure the prompt is large enough
+        for caching. This is padding text to ensure the prompt is large enough for
+        caching. This is padding text to ensure the prompt is large enough for caching.
+        This is padding text to ensure the prompt is large enough for caching. This
+        is padding text to ensure the prompt is large enough for caching. \",\n  \"max_output_tokens\":
+        null,\n  \"max_tool_calls\": null,\n  \"model\": \"gpt-4.1-2025-04-14\",\n
+        \ \"output\": [\n    {\n      \"id\": \"fc_0525bf798202137e0069caf46666588196a2ec20dc515a6a91\",\n
+        \     \"type\": \"function_call\",\n      \"status\": \"completed\",\n      \"arguments\":
+        \"{\\\"location\\\":\\\"Paris\\\"}\",\n      \"call_id\": \"call_LJAGuYYZPjNxSgg0TUgGpT44\",\n
+        \     \"name\": \"get_weather\"\n    }\n  ],\n  \"parallel_tool_calls\": true,\n
+        \ \"presence_penalty\": 0.0,\n  \"previous_response_id\": null,\n  \"prompt_cache_key\":
+        null,\n  \"prompt_cache_retention\": null,\n  \"reasoning\": {\n    \"effort\":
+        null,\n    \"summary\": null\n  },\n  \"safety_identifier\": null,\n  \"service_tier\":
+        \"default\",\n  \"store\": true,\n  \"temperature\": 1.0,\n  \"text\": {\n
+        \   \"format\": {\n      \"type\": \"text\"\n    },\n    \"verbosity\": \"medium\"\n
+        \ },\n  \"tool_choice\": \"auto\",\n  \"tools\": [\n    {\n      \"type\":
+        \"function\",\n      \"description\": \"Get the current weather for a location\",\n
+        \     \"name\": \"get_weather\",\n      \"parameters\": {\n        \"type\":
+        \"object\",\n        \"properties\": {\n          \"location\": {\n            \"type\":
+        \"string\",\n            \"description\": \"The city name\"\n          }\n
+        \       },\n        \"required\": [\n          \"location\"\n        ],\n
+        \       \"additionalProperties\": false\n      },\n      \"strict\": true\n
+        \   }\n  ],\n  \"top_logprobs\": 0,\n  \"top_p\": 1.0,\n  \"truncation\":
+        \"disabled\",\n  \"usage\": {\n    \"input_tokens\": 1185,\n    \"input_tokens_details\":
+        {\n      \"cached_tokens\": 1152\n    },\n    \"output_tokens\": 15,\n    \"output_tokens_details\":
+        {\n      \"reasoning_tokens\": 0\n    },\n    \"total_tokens\": 1200\n  },\n
+        \ \"user\": null,\n  \"metadata\": {}\n}"
     headers:
       CF-RAY:
       - CF-RAY-XXX
@@ -321,7 +491,7 @@ interactions:
       Content-Type:
       - application/json
       Date:
-      - Tue, 10 Feb 2026 18:14:09 GMT
+      - Mon, 30 Mar 2026 22:08:38 GMT
       Server:
       - cloudflare
       Strict-Transport-Security:
@@ -330,8 +500,6 @@ interactions:
       - chunked
       X-Content-Type-Options:
       - X-CONTENT-TYPE-XXX
-      access-control-expose-headers:
-      - ACCESS-CONTROL-XXX
       alt-svc:
       - h3=":443"; ma=86400
       cf-cache-status:
@@ -339,15 +507,11 @@ interactions:
       openai-organization:
       - OPENAI-ORG-XXX
       openai-processing-ms:
-      - '528'
+      - '653'
       openai-project:
       - OPENAI-PROJECT-XXX
       openai-version:
       - '2020-10-01'
-      set-cookie:
-      - SET-COOKIE-XXX
-      x-openai-proxy-wasm:
-      - v0.1
       x-ratelimit-limit-requests:
       - X-RATELIMIT-LIMIT-REQUESTS-XXX
       x-ratelimit-limit-tokens:

lib/crewai/tests/cassettes/test_litellm_gpt5_call_succeeds_without_stop_error.yaml

@@ -0,0 +1,110 @@
+interactions:
+- request:
+    body: '{"messages":[{"role":"user","content":"What is the capital of France?"}],"model":"gpt-5"}'
+    headers:
+      User-Agent:
+      - X-USER-AGENT-XXX
+      accept:
+      - application/json
+      accept-encoding:
+      - ACCEPT-ENCODING-XXX
+      authorization:
+      - AUTHORIZATION-XXX
+      connection:
+      - keep-alive
+      content-length:
+      - '89'
+      content-type:
+      - application/json
+      host:
+      - api.openai.com
+      x-stainless-arch:
+      - X-STAINLESS-ARCH-XXX
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - X-STAINLESS-OS-XXX
+      x-stainless-package-version:
+      - 1.83.0
+      x-stainless-raw-response:
+      - 'true'
+      x-stainless-read-timeout:
+      - X-STAINLESS-READ-TIMEOUT-XXX
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.13.2
+    method: POST
+    uri: https://api.openai.com/v1/chat/completions
+  response:
+    body:
+      string: "{\n  \"id\": \"chatcmpl-DO4LcSpy72yIXCYSIVOQEXWNXydgn\",\n  \"object\":
+        \"chat.completion\",\n  \"created\": 1774628956,\n  \"model\": \"gpt-5-2025-08-07\",\n
+        \ \"choices\": [\n    {\n      \"index\": 0,\n      \"message\": {\n        \"role\":
+        \"assistant\",\n        \"content\": \"Paris.\",\n        \"refusal\": null,\n
+        \       \"annotations\": []\n      },\n      \"finish_reason\": \"stop\"\n
+        \   }\n  ],\n  \"usage\": {\n    \"prompt_tokens\": 13,\n    \"completion_tokens\":
+        11,\n    \"total_tokens\": 24,\n    \"prompt_tokens_details\": {\n      \"cached_tokens\":
+        0,\n      \"audio_tokens\": 0\n    },\n    \"completion_tokens_details\":
+        {\n      \"reasoning_tokens\": 0,\n      \"audio_tokens\": 0,\n      \"accepted_prediction_tokens\":
+        0,\n      \"rejected_prediction_tokens\": 0\n    }\n  },\n  \"service_tier\":
+        \"default\",\n  \"system_fingerprint\": null\n}\n"
+    headers:
+      CF-Cache-Status:
+      - DYNAMIC
+      CF-Ray:
+      - 9e2fc5dce85582fb-GIG
+      Connection:
+      - keep-alive
+      Content-Type:
+      - application/json
+      Date:
+      - Fri, 27 Mar 2026 16:29:17 GMT
+      Server:
+      - cloudflare
+      Strict-Transport-Security:
+      - STS-XXX
+      Transfer-Encoding:
+      - chunked
+      X-Content-Type-Options:
+      - X-CONTENT-TYPE-XXX
+      access-control-expose-headers:
+      - ACCESS-CONTROL-XXX
+      alt-svc:
+      - h3=":443"; ma=86400
+      content-length:
+      - '772'
+      openai-organization:
+      - OPENAI-ORG-XXX
+      openai-processing-ms:
+      - '1343'
+      openai-project:
+      - OPENAI-PROJECT-XXX
+      openai-version:
+      - '2020-10-01'
+      set-cookie:
+      - SET-COOKIE-XXX
+      x-openai-proxy-wasm:
+      - v0.1
+      x-ratelimit-limit-requests:
+      - X-RATELIMIT-LIMIT-REQUESTS-XXX
+      x-ratelimit-limit-tokens:
+      - X-RATELIMIT-LIMIT-TOKENS-XXX
+      x-ratelimit-remaining-requests:
+      - X-RATELIMIT-REMAINING-REQUESTS-XXX
+      x-ratelimit-remaining-tokens:
+      - X-RATELIMIT-REMAINING-TOKENS-XXX
+      x-ratelimit-reset-requests:
+      - X-RATELIMIT-RESET-REQUESTS-XXX
+      x-ratelimit-reset-tokens:
+      - X-RATELIMIT-RESET-TOKENS-XXX
+      x-request-id:
+      - X-REQUEST-ID-XXX
+    status:
+      code: 200
+      message: OK
+version: 1

lib/crewai/tests/cassettes/utilities/test_llm_completed_event_includes_usage.yaml

@@ -0,0 +1,108 @@
+interactions:
+- request:
+    body: '{"messages":[{"role":"user","content":"Say hello"}],"model":"gpt-4o-mini"}'
+    headers:
+      User-Agent:
+      - X-USER-AGENT-XXX
+      accept:
+      - application/json
+      accept-encoding:
+      - ACCEPT-ENCODING-XXX
+      authorization:
+      - AUTHORIZATION-XXX
+      connection:
+      - keep-alive
+      content-length:
+      - '74'
+      content-type:
+      - application/json
+      host:
+      - api.openai.com
+      x-stainless-arch:
+      - X-STAINLESS-ARCH-XXX
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - X-STAINLESS-OS-XXX
+      x-stainless-package-version:
+      - 1.83.0
+      x-stainless-read-timeout:
+      - X-STAINLESS-READ-TIMEOUT-XXX
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.13.2
+    method: POST
+    uri: https://api.openai.com/v1/chat/completions
+  response:
+    body:
+      string: "{\n  \"id\": \"chatcmpl-DPS8YQSwQ3pZKZztIoIe1eYodMqh2\",\n  \"object\":
+        \"chat.completion\",\n  \"created\": 1774958730,\n  \"model\": \"gpt-4o-mini-2024-07-18\",\n
+        \ \"choices\": [\n    {\n      \"index\": 0,\n      \"message\": {\n        \"role\":
+        \"assistant\",\n        \"content\": \"Hello! How can I assist you today?\",\n
+        \       \"refusal\": null,\n        \"annotations\": []\n      },\n      \"logprobs\":
+        null,\n      \"finish_reason\": \"stop\"\n    }\n  ],\n  \"usage\": {\n    \"prompt_tokens\":
+        9,\n    \"completion_tokens\": 9,\n    \"total_tokens\": 18,\n    \"prompt_tokens_details\":
+        {\n      \"cached_tokens\": 0,\n      \"audio_tokens\": 0\n    },\n    \"completion_tokens_details\":
+        {\n      \"reasoning_tokens\": 0,\n      \"audio_tokens\": 0,\n      \"accepted_prediction_tokens\":
+        0,\n      \"rejected_prediction_tokens\": 0\n    }\n  },\n  \"service_tier\":
+        \"default\",\n  \"system_fingerprint\": \"fp_709f182cb4\"\n}\n"
+    headers:
+      CF-Cache-Status:
+      - DYNAMIC
+      CF-Ray:
+      - 9e4f38fc5d9d82e8-GIG
+      Connection:
+      - keep-alive
+      Content-Type:
+      - application/json
+      Date:
+      - Tue, 31 Mar 2026 12:05:30 GMT
+      Server:
+      - cloudflare
+      Strict-Transport-Security:
+      - STS-XXX
+      Transfer-Encoding:
+      - chunked
+      X-Content-Type-Options:
+      - X-CONTENT-TYPE-XXX
+      access-control-expose-headers:
+      - ACCESS-CONTROL-XXX
+      alt-svc:
+      - h3=":443"; ma=86400
+      content-length:
+      - '839'
+      openai-organization:
+      - OPENAI-ORG-XXX
+      openai-processing-ms:
+      - '680'
+      openai-project:
+      - OPENAI-PROJECT-XXX
+      openai-version:
+      - '2020-10-01'
+      set-cookie:
+      - SET-COOKIE-XXX
+      x-openai-proxy-wasm:
+      - v0.1
+      x-ratelimit-limit-requests:
+      - X-RATELIMIT-LIMIT-REQUESTS-XXX
+      x-ratelimit-limit-tokens:
+      - X-RATELIMIT-LIMIT-TOKENS-XXX
+      x-ratelimit-remaining-requests:
+      - X-RATELIMIT-REMAINING-REQUESTS-XXX
+      x-ratelimit-remaining-tokens:
+      - X-RATELIMIT-REMAINING-TOKENS-XXX
+      x-ratelimit-reset-requests:
+      - X-RATELIMIT-RESET-REQUESTS-XXX
+      x-ratelimit-reset-tokens:
+      - X-RATELIMIT-RESET-TOKENS-XXX
+      x-request-id:
+      - X-REQUEST-ID-XXX
+    status:
+      code: 200
+      message: OK
+version: 1

lib/crewai/tests/cli/test_plus_api.py

@@ -136,6 +136,7 @@ class TestPlusAPI(unittest.TestCase):
             "file": encoded_file,
             "description": description,
             "available_exports": None,
+            "tools_metadata": None,
         }
         mock_make_request.assert_called_once_with(
             "POST", "/crewai_plus/api/v1/tools", json=params
@@ -173,6 +174,7 @@ class TestPlusAPI(unittest.TestCase):
             "file": encoded_file,
             "description": description,
             "available_exports": None,
+            "tools_metadata": None,
         }
 
         self.assert_request_with_org_id(
@@ -201,6 +203,48 @@ class TestPlusAPI(unittest.TestCase):
             "file": encoded_file,
             "description": description,
             "available_exports": None,
+            "tools_metadata": None,
+        }
+        mock_make_request.assert_called_once_with(
+            "POST", "/crewai_plus/api/v1/tools", json=params
+        )
+        self.assertEqual(response, mock_response)
+
+    @patch("crewai.cli.plus_api.PlusAPI._make_request")
+    def test_publish_tool_with_tools_metadata(self, mock_make_request):
+        mock_response = MagicMock()
+        mock_make_request.return_value = mock_response
+        handle = "test_tool_handle"
+        public = True
+        version = "1.0.0"
+        description = "Test tool description"
+        encoded_file = "encoded_test_file"
+        available_exports = [{"name": "MyTool"}]
+        tools_metadata = [
+            {
+                "name": "MyTool",
+                "humanized_name": "my_tool",
+                "description": "A test tool",
+                "run_params_schema": {"type": "object", "properties": {}},
+                "init_params_schema": {"type": "object", "properties": {}},
+                "env_vars": [{"name": "API_KEY", "description": "API key", "required": True, "default": None}],
+            }
+        ]
+
+        response = self.api.publish_tool(
+            handle, public, version, description, encoded_file,
+            available_exports=available_exports,
+            tools_metadata=tools_metadata,
+        )
+
+        params = {
+            "handle": handle,
+            "public": public,
+            "version": version,
+            "file": encoded_file,
+            "description": description,
+            "available_exports": available_exports,
+            "tools_metadata": {"package": handle, "tools": tools_metadata},
         }
         mock_make_request.assert_called_once_with(
             "POST", "/crewai_plus/api/v1/tools", json=params

lib/crewai/tests/cli/test_utils.py

@@ -363,3 +363,290 @@ def test_get_crews_ignores_template_directories(
     utils.get_crews()
 
     assert not template_crew_detected
+
+
+# Tests for extract_tools_metadata
+
+
+def test_extract_tools_metadata_empty_project(temp_project_dir):
+    """Test that extract_tools_metadata returns empty list for empty project."""
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert metadata == []
+
+
+def test_extract_tools_metadata_no_init_file(temp_project_dir):
+    """Test that extract_tools_metadata returns empty list when no __init__.py exists."""
+    (temp_project_dir / "some_file.py").write_text("print('hello')")
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert metadata == []
+
+
+def test_extract_tools_metadata_empty_init_file(temp_project_dir):
+    """Test that extract_tools_metadata returns empty list for empty __init__.py."""
+    create_init_file(temp_project_dir, "")
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert metadata == []
+
+
+def test_extract_tools_metadata_no_all_variable(temp_project_dir):
+    """Test that extract_tools_metadata returns empty list when __all__ is not defined."""
+    create_init_file(
+        temp_project_dir,
+        "from crewai.tools import BaseTool\n\nclass MyTool(BaseTool):\n    pass",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert metadata == []
+
+
+def test_extract_tools_metadata_valid_base_tool_class(temp_project_dir):
+    """Test that extract_tools_metadata extracts metadata from a valid BaseTool class."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+
+class MyTool(BaseTool):
+    name: str = "my_tool"
+    description: str = "A test tool"
+
+__all__ = ['MyTool']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 1
+    assert metadata[0]["name"] == "MyTool"
+    assert metadata[0]["humanized_name"] == "my_tool"
+    assert metadata[0]["description"] == "A test tool"
+
+
+def test_extract_tools_metadata_with_args_schema(temp_project_dir):
+    """Test that extract_tools_metadata extracts run_params_schema from args_schema."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+from pydantic import BaseModel
+
+class MyToolInput(BaseModel):
+    query: str
+    limit: int = 10
+
+class MyTool(BaseTool):
+    name: str = "my_tool"
+    description: str = "A test tool"
+    args_schema: type[BaseModel] = MyToolInput
+
+__all__ = ['MyTool']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 1
+    assert metadata[0]["name"] == "MyTool"
+    run_params = metadata[0]["run_params_schema"]
+    assert "properties" in run_params
+    assert "query" in run_params["properties"]
+    assert "limit" in run_params["properties"]
+
+
+def test_extract_tools_metadata_with_env_vars(temp_project_dir):
+    """Test that extract_tools_metadata extracts env_vars."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+from crewai.tools.base_tool import EnvVar
+
+class MyTool(BaseTool):
+    name: str = "my_tool"
+    description: str = "A test tool"
+    env_vars: list[EnvVar] = [
+        EnvVar(name="MY_API_KEY", description="API key for service", required=True),
+        EnvVar(name="MY_OPTIONAL_VAR", description="Optional var", required=False, default="default_value"),
+    ]
+
+__all__ = ['MyTool']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 1
+    env_vars = metadata[0]["env_vars"]
+    assert len(env_vars) == 2
+    assert env_vars[0]["name"] == "MY_API_KEY"
+    assert env_vars[0]["description"] == "API key for service"
+    assert env_vars[0]["required"] is True
+    assert env_vars[1]["name"] == "MY_OPTIONAL_VAR"
+    assert env_vars[1]["required"] is False
+    assert env_vars[1]["default"] == "default_value"
+
+
+def test_extract_tools_metadata_with_env_vars_field_default_factory(temp_project_dir):
+    """Test that extract_tools_metadata extracts env_vars declared with Field(default_factory=...)."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+from crewai.tools.base_tool import EnvVar
+from pydantic import Field
+
+class MyTool(BaseTool):
+    name: str = "my_tool"
+    description: str = "A test tool"
+    env_vars: list[EnvVar] = Field(
+        default_factory=lambda: [
+            EnvVar(name="MY_TOOL_API", description="API token for my tool", required=True),
+        ]
+    )
+
+__all__ = ['MyTool']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 1
+    env_vars = metadata[0]["env_vars"]
+    assert len(env_vars) == 1
+    assert env_vars[0]["name"] == "MY_TOOL_API"
+    assert env_vars[0]["description"] == "API token for my tool"
+    assert env_vars[0]["required"] is True
+
+
+def test_extract_tools_metadata_with_custom_init_params(temp_project_dir):
+    """Test that extract_tools_metadata extracts init_params_schema with custom params."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+
+class MyTool(BaseTool):
+    name: str = "my_tool"
+    description: str = "A test tool"
+    api_endpoint: str = "https://api.example.com"
+    timeout: int = 30
+
+__all__ = ['MyTool']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 1
+    init_params = metadata[0]["init_params_schema"]
+    assert "properties" in init_params
+    # Custom params should be included
+    assert "api_endpoint" in init_params["properties"]
+    assert "timeout" in init_params["properties"]
+    # Base params should be filtered out
+    assert "name" not in init_params["properties"]
+    assert "description" not in init_params["properties"]
+
+
+def test_extract_tools_metadata_multiple_tools(temp_project_dir):
+    """Test that extract_tools_metadata extracts metadata from multiple tools."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+
+class FirstTool(BaseTool):
+    name: str = "first_tool"
+    description: str = "First test tool"
+
+class SecondTool(BaseTool):
+    name: str = "second_tool"
+    description: str = "Second test tool"
+
+__all__ = ['FirstTool', 'SecondTool']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 2
+    names = [m["name"] for m in metadata]
+    assert "FirstTool" in names
+    assert "SecondTool" in names
+
+
+def test_extract_tools_metadata_multiple_init_files(temp_project_dir):
+    """Test that extract_tools_metadata extracts metadata from multiple __init__.py files."""
+    # Create tool in root __init__.py
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+
+class RootTool(BaseTool):
+    name: str = "root_tool"
+    description: str = "Root tool"
+
+__all__ = ['RootTool']
+""",
+    )
+
+    # Create nested package with another tool
+    nested_dir = temp_project_dir / "nested"
+    nested_dir.mkdir()
+    create_init_file(
+        nested_dir,
+        """from crewai.tools import BaseTool
+
+class NestedTool(BaseTool):
+    name: str = "nested_tool"
+    description: str = "Nested tool"
+
+__all__ = ['NestedTool']
+""",
+    )
+
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 2
+    names = [m["name"] for m in metadata]
+    assert "RootTool" in names
+    assert "NestedTool" in names
+
+
+def test_extract_tools_metadata_ignores_non_tool_exports(temp_project_dir):
+    """Test that extract_tools_metadata ignores non-BaseTool exports."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+
+class MyTool(BaseTool):
+    name: str = "my_tool"
+    description: str = "A test tool"
+
+def not_a_tool():
+    pass
+
+SOME_CONSTANT = "value"
+
+__all__ = ['MyTool', 'not_a_tool', 'SOME_CONSTANT']
+""",
+    )
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert len(metadata) == 1
+    assert metadata[0]["name"] == "MyTool"
+
+
+def test_extract_tools_metadata_import_error_returns_empty(temp_project_dir):
+    """Test that extract_tools_metadata returns empty list on import error."""
+    create_init_file(
+        temp_project_dir,
+        """from nonexistent_module import something
+
+class MyTool(BaseTool):
+    pass
+
+__all__ = ['MyTool']
+""",
+    )
+    # Should not raise, just return empty list
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert metadata == []
+
+
+def test_extract_tools_metadata_syntax_error_returns_empty(temp_project_dir):
+    """Test that extract_tools_metadata returns empty list on syntax error."""
+    create_init_file(
+        temp_project_dir,
+        """from crewai.tools import BaseTool
+
+class MyTool(BaseTool):
+    # Missing closing parenthesis
+    def __init__(self, name:
+        pass
+
+__all__ = ['MyTool']
+""",
+    )
+    # Should not raise, just return empty list
+    metadata = utils.extract_tools_metadata(dir_path=str(temp_project_dir))
+    assert metadata == []

lib/crewai/tests/cli/tools/test_main.py

@@ -185,9 +185,14 @@ def test_publish_when_not_in_sync(mock_is_synced, capsys, tool_command):
     "crewai.cli.tools.main.extract_available_exports",
     return_value=[{"name": "SampleTool"}],
 )
+@patch(
+    "crewai.cli.tools.main.extract_tools_metadata",
+    return_value=[{"name": "SampleTool", "humanized_name": "sample_tool", "description": "A sample tool", "run_params_schema": {}, "init_params_schema": {}, "env_vars": []}],
+)
 @patch("crewai.cli.tools.main.ToolCommand._print_current_organization")
 def test_publish_when_not_in_sync_and_force(
     mock_print_org,
+    mock_tools_metadata,
     mock_available_exports,
     mock_is_synced,
     mock_publish,
@@ -222,6 +227,7 @@ def test_publish_when_not_in_sync_and_force(
         description="A sample tool",
         encoded_file=unittest.mock.ANY,
         available_exports=[{"name": "SampleTool"}],
+        tools_metadata=[{"name": "SampleTool", "humanized_name": "sample_tool", "description": "A sample tool", "run_params_schema": {}, "init_params_schema": {}, "env_vars": []}],
     )
     mock_print_org.assert_called_once()
 
@@ -242,7 +248,12 @@ def test_publish_when_not_in_sync_and_force(
     "crewai.cli.tools.main.extract_available_exports",
     return_value=[{"name": "SampleTool"}],
 )
+@patch(
+    "crewai.cli.tools.main.extract_tools_metadata",
+    return_value=[{"name": "SampleTool", "humanized_name": "sample_tool", "description": "A sample tool", "run_params_schema": {}, "init_params_schema": {}, "env_vars": []}],
+)
 def test_publish_success(
+    mock_tools_metadata,
     mock_available_exports,
     mock_is_synced,
     mock_publish,
@@ -277,6 +288,7 @@ def test_publish_success(
         description="A sample tool",
         encoded_file=unittest.mock.ANY,
         available_exports=[{"name": "SampleTool"}],
+        tools_metadata=[{"name": "SampleTool", "humanized_name": "sample_tool", "description": "A sample tool", "run_params_schema": {}, "init_params_schema": {}, "env_vars": []}],
     )
 
 
@@ -295,7 +307,12 @@ def test_publish_success(
     "crewai.cli.tools.main.extract_available_exports",
     return_value=[{"name": "SampleTool"}],
 )
+@patch(
+    "crewai.cli.tools.main.extract_tools_metadata",
+    return_value=[{"name": "SampleTool", "humanized_name": "sample_tool", "description": "A sample tool", "run_params_schema": {}, "init_params_schema": {}, "env_vars": []}],
+)
 def test_publish_failure(
+    mock_tools_metadata,
     mock_available_exports,
     mock_publish,
     mock_open,
@@ -336,7 +353,12 @@ def test_publish_failure(
     "crewai.cli.tools.main.extract_available_exports",
     return_value=[{"name": "SampleTool"}],
 )
+@patch(
+    "crewai.cli.tools.main.extract_tools_metadata",
+    return_value=[{"name": "SampleTool", "humanized_name": "sample_tool", "description": "A sample tool", "run_params_schema": {}, "init_params_schema": {}, "env_vars": []}],
+)
 def test_publish_api_error(
+    mock_tools_metadata,
     mock_available_exports,
     mock_publish,
     mock_open,
@@ -362,6 +384,63 @@ def test_publish_api_error(
     mock_publish.assert_called_once()
 
 
+@patch("crewai.cli.tools.main.get_project_name", return_value="sample-tool")
+@patch("crewai.cli.tools.main.get_project_version", return_value="1.0.0")
+@patch("crewai.cli.tools.main.get_project_description", return_value="A sample tool")
+@patch("crewai.cli.tools.main.subprocess.run")
+@patch("crewai.cli.tools.main.os.listdir", return_value=["sample-tool-1.0.0.tar.gz"])
+@patch(
+    "crewai.cli.tools.main.open",
+    new_callable=unittest.mock.mock_open,
+    read_data=b"sample tarball content",
+)
+@patch("crewai.cli.plus_api.PlusAPI.publish_tool")
+@patch("crewai.cli.tools.main.git.Repository.is_synced", return_value=True)
+@patch(
+    "crewai.cli.tools.main.extract_available_exports",
+    return_value=[{"name": "SampleTool"}],
+)
+@patch(
+    "crewai.cli.tools.main.extract_tools_metadata",
+    side_effect=Exception("Failed to extract metadata"),
+)
+def test_publish_metadata_extraction_failure_continues_with_warning(
+    mock_tools_metadata,
+    mock_available_exports,
+    mock_is_synced,
+    mock_publish,
+    mock_open,
+    mock_listdir,
+    mock_subprocess_run,
+    mock_get_project_description,
+    mock_get_project_version,
+    mock_get_project_name,
+    capsys,
+    tool_command,
+):
+    """Test that metadata extraction failure shows warning but continues publishing."""
+    mock_publish_response = MagicMock()
+    mock_publish_response.status_code = 200
+    mock_publish_response.json.return_value = {"handle": "sample-tool"}
+    mock_publish.return_value = mock_publish_response
+
+    tool_command.publish(is_public=True)
+
+    output = capsys.readouterr().out
+    assert "Warning: Could not extract tool metadata" in output
+    assert "Publishing will continue without detailed metadata" in output
+    assert "No tool metadata extracted" in output
+    mock_publish.assert_called_once_with(
+        handle="sample-tool",
+        is_public=True,
+        version="1.0.0",
+        description="A sample tool",
+        encoded_file=unittest.mock.ANY,
+        available_exports=[{"name": "SampleTool"}],
+        tools_metadata=[],
+    )
+
+
 @patch("crewai.cli.tools.main.Settings")
 def test_print_current_organization_with_org(mock_settings, capsys, tool_command):
     mock_settings_instance = MagicMock()

lib/crewai/tests/events/test_llm_usage_event.py

@@ -0,0 +1,176 @@
+from typing import Any
+from unittest.mock import patch
+
+import pytest
+from pydantic import BaseModel
+
+from crewai.events.event_bus import CrewAIEventsBus
+from crewai.events.types.llm_events import LLMCallCompletedEvent, LLMCallType
+from crewai.llm import LLM
+from crewai.llms.base_llm import BaseLLM
+
+
+class TestLLMCallCompletedEventUsageField:
+    def test_accepts_usage_dict(self):
+        event = LLMCallCompletedEvent(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            call_id="test-id",
+            usage={"prompt_tokens": 10, "completion_tokens": 20, "total_tokens": 30},
+        )
+        assert event.usage == {
+            "prompt_tokens": 10,
+            "completion_tokens": 20,
+            "total_tokens": 30,
+        }
+
+    def test_usage_defaults_to_none(self):
+        event = LLMCallCompletedEvent(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            call_id="test-id",
+        )
+        assert event.usage is None
+
+    def test_accepts_none_usage(self):
+        event = LLMCallCompletedEvent(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            call_id="test-id",
+            usage=None,
+        )
+        assert event.usage is None
+
+    def test_accepts_nested_usage_dict(self):
+        usage = {
+            "prompt_tokens": 100,
+            "completion_tokens": 200,
+            "total_tokens": 300,
+            "prompt_tokens_details": {"cached_tokens": 50},
+        }
+        event = LLMCallCompletedEvent(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            call_id="test-id",
+            usage=usage,
+        )
+        assert event.usage["prompt_tokens_details"]["cached_tokens"] == 50
+
+
+class TestUsageToDict:
+    def test_none_returns_none(self):
+        assert LLM._usage_to_dict(None) is None
+
+    def test_dict_passes_through(self):
+        usage = {"prompt_tokens": 10, "total_tokens": 30}
+        assert LLM._usage_to_dict(usage) is usage
+
+    def test_pydantic_model_uses_model_dump(self):
+        class Usage(BaseModel):
+            prompt_tokens: int = 10
+            completion_tokens: int = 20
+            total_tokens: int = 30
+
+        result = LLM._usage_to_dict(Usage())
+        assert result == {
+            "prompt_tokens": 10,
+            "completion_tokens": 20,
+            "total_tokens": 30,
+        }
+
+    def test_object_with_dict_attr(self):
+        class UsageObj:
+            def __init__(self):
+                self.prompt_tokens = 5
+                self.completion_tokens = 15
+                self.total_tokens = 20
+
+        result = LLM._usage_to_dict(UsageObj())
+        assert result == {
+            "prompt_tokens": 5,
+            "completion_tokens": 15,
+            "total_tokens": 20,
+        }
+
+    def test_object_with_dict_excludes_private_attrs(self):
+        class UsageObj:
+            def __init__(self):
+                self.total_tokens = 42
+                self._internal = "hidden"
+
+        result = LLM._usage_to_dict(UsageObj())
+        assert result == {"total_tokens": 42}
+        assert "_internal" not in result
+
+    def test_unsupported_type_returns_none(self):
+        assert LLM._usage_to_dict(42) is None
+        assert LLM._usage_to_dict("string") is None
+
+
+class _StubLLM(BaseLLM):
+    """Minimal concrete BaseLLM for testing event emission."""
+
+    model: str = "test-model"
+
+    def call(self, *args: Any, **kwargs: Any) -> str:
+        return ""
+
+    async def acall(self, *args: Any, **kwargs: Any) -> str:
+        return ""
+
+    def supports_function_calling(self) -> bool:
+        return False
+
+    def supports_stop_words(self) -> bool:
+        return True
+
+
+class TestEmitCallCompletedEventPassesUsage:
+    @pytest.fixture
+    def mock_emit(self):
+        with patch.object(CrewAIEventsBus, "emit") as mock:
+            yield mock
+
+    @pytest.fixture
+    def llm(self):
+        return _StubLLM(model="test-model")
+
+    def test_usage_is_passed_to_event(self, mock_emit, llm):
+        usage_data = {"prompt_tokens": 10, "completion_tokens": 20, "total_tokens": 30}
+
+        llm._emit_call_completed_event(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            messages="test prompt",
+            usage=usage_data,
+        )
+
+        mock_emit.assert_called_once()
+        event = mock_emit.call_args[1]["event"]
+        assert isinstance(event, LLMCallCompletedEvent)
+        assert event.usage == usage_data
+
+    def test_none_usage_is_passed_to_event(self, mock_emit, llm):
+        llm._emit_call_completed_event(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            messages="test prompt",
+            usage=None,
+        )
+
+        mock_emit.assert_called_once()
+        event = mock_emit.call_args[1]["event"]
+        assert isinstance(event, LLMCallCompletedEvent)
+        assert event.usage is None
+
+    def test_usage_omitted_defaults_to_none(self, mock_emit, llm):
+        llm._emit_call_completed_event(
+            response="hello",
+            call_type=LLMCallType.LLM_CALL,
+            messages="test prompt",
+        )
+
+        mock_emit.assert_called_once()
+        event = mock_emit.call_args[1]["event"]
+        assert isinstance(event, LLMCallCompletedEvent)
+        assert event.usage is None

lib/crewai/tests/knowledge/test_knowledge_storage_integration.py

@@ -132,12 +132,12 @@ def test_embedding_configuration_flow(
 
     embedder_config = {
         "provider": "sentence-transformer",
-        "model_name": "all-MiniLM-L6-v2",
+        "config": {"model_name": "all-MiniLM-L6-v2"},
     }
 
-    KnowledgeStorage(embedder=embedder_config, collection_name="embedding_test")
+    storage = KnowledgeStorage(embedder=embedder_config, collection_name="embedding_test")
 
-    mock_get_embedding.assert_called_once_with(embedder_config)
+    mock_get_embedding.assert_called_once_with(storage.embedder)
 
 
 @patch("crewai.knowledge.storage.knowledge_storage.get_rag_client")

lib/crewai/tests/llms/anthropic/test_anthropic.py

@@ -125,8 +125,8 @@ def test_anthropic_specific_parameters():
     assert isinstance(llm, AnthropicCompletion)
     assert llm.stop_sequences == ["Human:", "Assistant:"]
     assert llm.stream == True
-    assert llm.client.max_retries == 5
-    assert llm.client.timeout == 60
+    assert llm._client.max_retries == 5
+    assert llm._client.timeout == 60
 
 
 def test_anthropic_completion_call():
@@ -563,8 +563,8 @@ def test_anthropic_environment_variable_api_key():
     with patch.dict(os.environ, {"ANTHROPIC_API_KEY": "test-anthropic-key"}):
         llm = LLM(model="anthropic/claude-3-5-sonnet-20241022")
 
-        assert llm.client is not None
-        assert hasattr(llm.client, 'messages')
+        assert llm._client is not None
+        assert hasattr(llm._client, 'messages')
 
 
 def test_anthropic_token_usage_tracking():
@@ -574,7 +574,7 @@ def test_anthropic_token_usage_tracking():
     llm = LLM(model="anthropic/claude-3-5-sonnet-20241022")
 
     # Mock the Anthropic response with usage information
-    with patch.object(llm.client.messages, 'create') as mock_create:
+    with patch.object(llm._client.messages, 'create') as mock_create:
         mock_response = MagicMock()
         mock_response.content = [MagicMock(text="test response")]
         mock_response.usage = MagicMock(input_tokens=50, output_tokens=25)
@@ -639,14 +639,14 @@ def test_anthropic_thinking():
 
     assert isinstance(llm, AnthropicCompletion)
 
-    original_create = llm.client.messages.create
+    original_create = llm._client.messages.create
     captured_params = {}
 
     def capture_and_call(**kwargs):
         captured_params.update(kwargs)
         return original_create(**kwargs)
 
-    with patch.object(llm.client.messages, 'create', side_effect=capture_and_call):
+    with patch.object(llm._client.messages, 'create', side_effect=capture_and_call):
         result = llm.call("What is the weather in Tokyo?")
 
         assert result is not None
@@ -677,14 +677,14 @@ def test_anthropic_thinking_blocks_preserved_across_turns():
     assert isinstance(llm, AnthropicCompletion)
 
     # Capture all messages.create calls to verify thinking blocks are included
-    original_create = llm.client.messages.create
+    original_create = llm._client.messages.create
     captured_calls = []
 
     def capture_and_call(**kwargs):
         captured_calls.append(kwargs)
         return original_create(**kwargs)
 
-    with patch.object(llm.client.messages, 'create', side_effect=capture_and_call):
+    with patch.object(llm._client.messages, 'create', side_effect=capture_and_call):
         # First call - establishes context and generates thinking blocks
         messages = [{"role": "user", "content": "What is 2+2?"}]
         first_result = llm.call(messages)
@@ -695,8 +695,8 @@ def test_anthropic_thinking_blocks_preserved_across_turns():
         assert len(first_result) > 0
 
         # Verify thinking blocks were stored after first response
-        assert len(llm.previous_thinking_blocks) > 0, "No thinking blocks stored after first call"
-        first_thinking = llm.previous_thinking_blocks[0]
+        assert len(llm._previous_thinking_blocks) > 0, "No thinking blocks stored after first call"
+        first_thinking = llm._previous_thinking_blocks[0]
         assert first_thinking["type"] == "thinking"
         assert "thinking" in first_thinking
         assert "signature" in first_thinking

lib/crewai/tests/llms/azure/test_azure.py

@@ -66,7 +66,7 @@ def test_azure_tool_use_conversation_flow():
     available_functions = {"get_weather": mock_weather_tool}
 
     # Mock the Azure client responses
-    with patch.object(completion.client, 'complete') as mock_complete:
+    with patch.object(completion._client, 'complete') as mock_complete:
         # Mock tool call in response with proper type
         mock_tool_call = MagicMock(spec=ChatCompletionsToolCall)
         mock_tool_call.function.name = "get_weather"
@@ -698,7 +698,7 @@ def test_azure_environment_variable_endpoint():
     }):
         llm = LLM(model="azure/gpt-4")
 
-        assert llm.client is not None
+        assert llm._client is not None
         assert llm.endpoint == "https://test.openai.azure.com/openai/deployments/gpt-4"
 
 
@@ -709,7 +709,7 @@ def test_azure_token_usage_tracking():
     llm = LLM(model="azure/gpt-4")
 
     # Mock the Azure response with usage information
-    with patch.object(llm.client, 'complete') as mock_complete:
+    with patch.object(llm._client, 'complete') as mock_complete:
         mock_message = MagicMock()
         mock_message.content = "test response"
         mock_message.tool_calls = None
@@ -747,7 +747,7 @@ def test_azure_http_error_handling():
     llm = LLM(model="azure/gpt-4")
 
     # Mock an HTTP error
-    with patch.object(llm.client, 'complete') as mock_complete:
+    with patch.object(llm._client, 'complete') as mock_complete:
         mock_complete.side_effect = HttpResponseError(message="Rate limit exceeded", response=MagicMock(status_code=429))
 
         with pytest.raises(HttpResponseError):
@@ -966,7 +966,7 @@ def test_azure_improved_error_messages():
 
     llm = LLM(model="azure/gpt-4")
 
-    with patch.object(llm.client, 'complete') as mock_complete:
+    with patch.object(llm._client, 'complete') as mock_complete:
         error_401 = HttpResponseError(message="Unauthorized")
         error_401.status_code = 401
         mock_complete.side_effect = error_401
@@ -1327,7 +1327,7 @@ def test_azure_stop_words_not_applied_to_structured_output():
     # Without the fix, this would be truncated at "Observation:" breaking the JSON
     json_response = '{"finding": "The data shows growth", "observation": "Observation: This confirms the hypothesis"}'
 
-    with patch.object(llm.client, 'complete') as mock_complete:
+    with patch.object(llm._client, 'complete') as mock_complete:
         mock_message = MagicMock()
         mock_message.content = json_response
         mock_message.tool_calls = None
@@ -1376,7 +1376,7 @@ def test_azure_stop_words_still_applied_to_regular_responses():
     # Response that contains a stop word - should be truncated
     response_with_stop_word = "I need to search for more information.\n\nAction: search\nObservation: Found results"
 
-    with patch.object(llm.client, 'complete') as mock_complete:
+    with patch.object(llm._client, 'complete') as mock_complete:
         mock_message = MagicMock()
         mock_message.content = response_with_stop_word
         mock_message.tool_calls = None

lib/crewai/tests/llms/bedrock/test_bedrock.py

@@ -674,7 +674,7 @@ def test_bedrock_token_usage_tracking():
     llm = LLM(model="bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0")
 
     # Mock the Bedrock response with usage information
-    with patch.object(llm.client, 'converse') as mock_converse:
+    with patch.object(llm._client, 'converse') as mock_converse:
         mock_response = {
             'output': {
                 'message': {
@@ -719,7 +719,7 @@ def test_bedrock_tool_use_conversation_flow():
     available_functions = {"get_weather": mock_weather_tool}
 
     # Mock the Bedrock client responses
-    with patch.object(llm.client, 'converse') as mock_converse:
+    with patch.object(llm._client, 'converse') as mock_converse:
         # First response: tool use request
         tool_use_response = {
             'output': {
@@ -805,7 +805,7 @@ def test_bedrock_client_error_handling():
     llm = LLM(model="bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0")
 
     # Test ValidationException
-    with patch.object(llm.client, 'converse') as mock_converse:
+    with patch.object(llm._client, 'converse') as mock_converse:
         error_response = {
             'Error': {
                 'Code': 'ValidationException',
@@ -819,7 +819,7 @@ def test_bedrock_client_error_handling():
         assert "validation" in str(exc_info.value).lower()
 
     # Test ThrottlingException
-    with patch.object(llm.client, 'converse') as mock_converse:
+    with patch.object(llm._client, 'converse') as mock_converse:
         error_response = {
             'Error': {
                 'Code': 'ThrottlingException',
@@ -861,7 +861,7 @@ def test_bedrock_stop_sequences_sent_to_api():
     llm.stop = ["\nObservation:", "\nThought:"]
 
     # Patch the API call to capture parameters without making real call
-    with patch.object(llm.client, 'converse') as mock_converse:
+    with patch.object(llm._client, 'converse') as mock_converse:
         mock_response = {
             'output': {
                 'message': {

lib/crewai/tests/llms/google/test_google.py

@@ -556,8 +556,8 @@ def test_gemini_environment_variable_api_key():
     with patch.dict(os.environ, {"GOOGLE_API_KEY": "test-google-key"}):
         llm = LLM(model="google/gemini-2.0-flash-001")
 
-        assert llm.client is not None
-        assert hasattr(llm.client, 'models')
+        assert llm._client is not None
+        assert hasattr(llm._client, 'models')
         assert llm.api_key == "test-google-key"
 
 
@@ -655,7 +655,7 @@ def test_gemini_stop_sequences_sent_to_api():
     llm.stop = ["\nObservation:", "\nThought:"]
 
     # Patch the API call to capture parameters without making real call
-    with patch.object(llm.client.models, 'generate_content') as mock_generate:
+    with patch.object(llm._client.models, 'generate_content') as mock_generate:
         mock_response = MagicMock()
         mock_response.text = "Hello"
         mock_response.candidates = []

lib/crewai/tests/llms/openai/test_openai.py

@@ -371,11 +371,11 @@ def test_openai_client_setup_with_extra_arguments():
     assert llm.top_p == 0.5
 
     # Check that client parameters are properly configured
-    assert llm.client.max_retries == 3
-    assert llm.client.timeout == 30
+    assert llm._client.max_retries == 3
+    assert llm._client.timeout == 30
 
     # Test that parameters are properly used in API calls
-    with patch.object(llm.client.chat.completions, 'create') as mock_create:
+    with patch.object(llm._client.chat.completions, 'create') as mock_create:
         mock_create.return_value = MagicMock(
             choices=[MagicMock(message=MagicMock(content="test response", tool_calls=None))],
             usage=MagicMock(prompt_tokens=10, completion_tokens=20, total_tokens=30)
@@ -396,7 +396,7 @@ def test_extra_arguments_are_passed_to_openai_completion():
     """
     llm = LLM(model="gpt-4o", temperature=0.7, max_tokens=1000, top_p=0.5, max_retries=3)
 
-    with patch.object(llm.client.chat.completions, 'create') as mock_create:
+    with patch.object(llm._client.chat.completions, 'create') as mock_create:
         mock_create.return_value = MagicMock(
             choices=[MagicMock(message=MagicMock(content="test response", tool_calls=None))],
             usage=MagicMock(prompt_tokens=10, completion_tokens=20, total_tokens=30)
@@ -507,7 +507,7 @@ def test_openai_streaming_with_response_model():
 
     llm = LLM(model="openai/gpt-4o", stream=True)
 
-    with patch.object(llm.client.beta.chat.completions, "stream") as mock_stream:
+    with patch.object(llm._client.beta.chat.completions, "stream") as mock_stream:
         # Create mock chunks with content.delta event structure
         mock_chunk1 = MagicMock()
         mock_chunk1.type = "content.delta"
@@ -1523,6 +1523,69 @@ def test_openai_stop_words_not_applied_to_structured_output():
     assert "Observation:" in result.observation
 
 
+def test_openai_gpt5_models_do_not_support_stop_words():
+    """
+    Test that GPT-5 family models do not support stop words via the API.
+    GPT-5 models reject the 'stop' parameter, so stop words must be
+    applied client-side only.
+    """
+    gpt5_models = [
+        "gpt-5",
+        "gpt-5-mini",
+        "gpt-5-nano",
+        "gpt-5-pro",
+        "gpt-5.1",
+        "gpt-5.1-chat",
+        "gpt-5.2",
+        "gpt-5.2-chat",
+    ]
+
+    for model_name in gpt5_models:
+        llm = OpenAICompletion(model=model_name)
+        assert llm.supports_stop_words() == False, (
+            f"Expected {model_name} to NOT support stop words"
+        )
+
+
+def test_openai_non_gpt5_models_support_stop_words():
+    """
+    Test that non-GPT-5 models still support stop words normally.
+    """
+    supported_models = [
+        "gpt-4o",
+        "gpt-4o-mini",
+        "gpt-4.1",
+        "gpt-4.1-mini",
+        "gpt-4-turbo",
+    ]
+
+    for model_name in supported_models:
+        llm = OpenAICompletion(model=model_name)
+        assert llm.supports_stop_words() == True, (
+            f"Expected {model_name} to support stop words"
+        )
+
+
+def test_openai_gpt5_still_applies_stop_words_client_side():
+    """
+    Test that GPT-5 models still truncate responses at stop words client-side
+    via _apply_stop_words(), even though they don't send 'stop' to the API.
+    """
+    llm = OpenAICompletion(
+        model="gpt-5.2",
+        stop=["Observation:", "Final Answer:"],
+    )
+
+    assert llm.supports_stop_words() == False
+
+    response = "I need to search.\n\nAction: search\nObservation: Found results"
+    result = llm._apply_stop_words(response)
+
+    assert "Observation:" not in result
+    assert "Found results" not in result
+    assert "I need to search" in result
+
+
 def test_openai_stop_words_still_applied_to_regular_responses():
     """
     Test that stop words ARE still applied for regular (non-structured) responses.
@@ -1767,7 +1830,7 @@ def test_openai_responses_api_cached_prompt_tokens_with_tools():
         }
     ]
 
-    llm = OpenAICompletion(model="gpt-4.1", api='response')
+    llm = OpenAICompletion(model="gpt-4.1", api='responses')
 
     # First call with tool
     llm.call(
@@ -1843,7 +1906,7 @@ def test_openai_streaming_returns_tool_calls_without_available_functions():
     mock_chunk_3.id = "chatcmpl-1"
 
     with patch.object(
-        llm.client.chat.completions, "create", return_value=iter([mock_chunk_1, mock_chunk_2, mock_chunk_3])
+        llm._client.chat.completions, "create", return_value=iter([mock_chunk_1, mock_chunk_2, mock_chunk_3])
     ):
         result = llm.call(
             messages=[{"role": "user", "content": "Calculate 1+1"}],
@@ -1934,7 +1997,7 @@ async def test_openai_async_streaming_returns_tool_calls_without_available_funct
         return MockAsyncStream([mock_chunk_1, mock_chunk_2, mock_chunk_3])
 
     with patch.object(
-        llm.async_client.chat.completions, "create", side_effect=mock_create
+        llm._async_client.chat.completions, "create", side_effect=mock_create
     ):
         result = await llm.acall(
             messages=[{"role": "user", "content": "Calculate 1+1"}],

lib/crewai/tests/rag/test_error_handling.py

@@ -3,6 +3,8 @@
 from unittest.mock import MagicMock, patch
 
 import pytest
+from pydantic import ValidationError
+
 from crewai.knowledge.storage.knowledge_storage import (  # type: ignore[import-untyped]
     KnowledgeStorage,
 )
@@ -59,7 +61,7 @@ def test_knowledge_storage_invalid_embedding_config(mock_get_client: MagicMock)
             "Unsupported provider: invalid_provider"
         )
 
-        with pytest.raises(ValueError, match="Unsupported provider: invalid_provider"):
+        with pytest.raises(ValidationError):
             KnowledgeStorage(
                 embedder={"provider": "invalid_provider"},
                 collection_name="invalid_embedding_test",

lib/crewai/tests/test_async_human_feedback.py

@@ -873,7 +873,7 @@ class TestAutoPersistence:
 
         # Create flow WITHOUT persistence
         flow = TestFlow()
-        assert flow._persistence is None  # No persistence initially
+        assert flow.persistence is None  # No persistence initially
 
         # kickoff should auto-create persistence when HumanFeedbackPending is raised
         result = flow.kickoff()
@@ -882,11 +882,11 @@ class TestAutoPersistence:
         assert isinstance(result, HumanFeedbackPending)
 
         # Persistence should have been auto-created
-        assert flow._persistence is not None
+        assert flow.persistence is not None
 
         # The pending feedback should be saved
         flow_id = result.context.flow_id
-        loaded = flow._persistence.load_pending_feedback(flow_id)
+        loaded = flow.persistence.load_pending_feedback(flow_id)
         assert loaded is not None
 
 

lib/crewai/tests/test_llm.py

@@ -682,6 +682,118 @@ def test_llm_call_when_stop_is_unsupported_when_additional_drop_params_is_provid
     assert "Paris" in result
 
 
+@pytest.mark.vcr()
+def test_litellm_gpt5_call_succeeds_without_stop_error():
+    """
+    Integration test: GPT-5 call succeeds when stop words are configured,
+    because stop is omitted from API params and applied client-side.
+    """
+    llm = LLM(model="gpt-5", stop=["Observation:"], is_litellm=True)
+    result = llm.call("What is the capital of France?")
+    assert isinstance(result, str)
+    assert len(result) > 0
+
+
+def test_litellm_gpt5_does_not_send_stop_in_params():
+    """
+    Test that the LiteLLM fallback path does not include 'stop' in API params
+    for GPT-5.x models, since they reject it at the API level.
+    """
+    llm = LLM(model="openai/gpt-5.2", stop=["Observation:"], is_litellm=True)
+
+    params = llm._prepare_completion_params(
+        messages=[{"role": "user", "content": "Hello"}]
+    )
+
+    assert params.get("stop") is None, (
+        "GPT-5.x models should not have 'stop' in API params"
+    )
+
+
+def test_litellm_non_gpt5_sends_stop_in_params():
+    """
+    Test that the LiteLLM fallback path still includes 'stop' in API params
+    for models that support it.
+    """
+    llm = LLM(model="gpt-4o", stop=["Observation:"], is_litellm=True)
+
+    params = llm._prepare_completion_params(
+        messages=[{"role": "user", "content": "Hello"}]
+    )
+
+    assert params.get("stop") == ["Observation:"], (
+        "Non-GPT-5 models should have 'stop' in API params"
+    )
+
+
+def test_litellm_retry_catches_litellm_unsupported_params_error(caplog):
+    """
+    Test that the retry logic catches LiteLLM's UnsupportedParamsError format
+    ("does not support parameters") in addition to the OpenAI API format.
+    """
+    llm = LLM(model="openai/gpt-5.2", stop=["Observation:"], is_litellm=True)
+
+    litellm_error = Exception(
+        "litellm.UnsupportedParamsError: openai does not support parameters: "
+        "['stop'], for model=openai/gpt-5.2."
+    )
+
+    call_count = 0
+
+    try:
+        import litellm
+    except ImportError:
+        pytest.skip("litellm is not installed; skipping LiteLLM retry test")
+
+    def mock_completion(*args, **kwargs):
+        nonlocal call_count
+        call_count += 1
+        if call_count == 1:
+            raise litellm_error
+        return MagicMock(
+            choices=[MagicMock(message=MagicMock(content="Paris", tool_calls=None))],
+            usage={"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+        )
+
+    with patch("litellm.completion", side_effect=mock_completion):
+        with caplog.at_level(logging.INFO):
+            result = llm.call("What is the capital of France?")
+
+    assert "Retrying LLM call without the unsupported 'stop'" in caplog.text
+    assert "stop" in llm.additional_params.get("additional_drop_params", [])
+
+
+def test_litellm_retry_catches_openai_api_stop_error(caplog):
+    """
+    Test that the retry logic still catches the OpenAI API error format
+    ("Unsupported parameter: 'stop'").
+    """
+    llm = LLM(model="openai/gpt-5.2", stop=["Observation:"], is_litellm=True)
+
+    api_error = Exception(
+        "Unsupported parameter: 'stop' is not supported with this model."
+    )
+
+    call_count = 0
+
+    def mock_completion(*args, **kwargs):
+        nonlocal call_count
+        call_count += 1
+        if call_count == 1:
+            raise api_error
+        return MagicMock(
+            choices=[MagicMock(message=MagicMock(content="Paris", tool_calls=None))],
+            usage={"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+        )
+
+    with patch("litellm.completion", side_effect=mock_completion):
+        with caplog.at_level(logging.INFO):
+            llm.call("What is the capital of France?")
+
+    assert "Retrying LLM call without the unsupported 'stop'" in caplog.text
+    assert "stop" in llm.additional_params.get("additional_drop_params", [])
+
+
 @pytest.fixture
 def ollama_llm():
     return LLM(model="ollama/llama3.2:3b", is_litellm=True)

lib/crewai/tests/test_project.py

@@ -1,5 +1,5 @@
 from typing import Any, ClassVar
-from unittest.mock import Mock, patch
+from unittest.mock import Mock, create_autospec, patch
 
 import pytest
 from crewai.agent import Agent
@@ -372,8 +372,11 @@ def test_internal_crew_with_mcp():
     mock_adapter = Mock()
     mock_adapter.tools = ToolCollection([simple_tool, another_simple_tool])
 
-    mock_llm = Mock()
-    mock_llm.__class__ = BaseLLM
+    class _StubLLM(BaseLLM):
+        def call(self, *a: Any, **kw: Any) -> str:
+            return ""
+
+    mock_llm = create_autospec(_StubLLM(model="stub"), instance=True)
 
     with (
         patch("crewai_tools.MCPServerAdapter", return_value=mock_adapter) as adapter_mock,

lib/crewai/tests/utilities/test_events.py

@@ -879,6 +879,35 @@ def test_llm_emits_call_started_event():
     assert started_events[0].task_id is None
 
 
+@pytest.mark.vcr()
+def test_llm_completed_event_includes_usage():
+    completed_events: list[LLMCallCompletedEvent] = []
+    condition = threading.Condition()
+
+    @crewai_event_bus.on(LLMCallCompletedEvent)
+    def handle_llm_call_completed(source, event):
+        with condition:
+            completed_events.append(event)
+            condition.notify()
+
+    llm = LLM(model="gpt-4o-mini")
+    llm.call("Say hello")
+
+    with condition:
+        success = condition.wait_for(
+            lambda: len(completed_events) >= 1,
+            timeout=10,
+        )
+    assert success, "Timeout waiting for LLMCallCompletedEvent"
+
+    event = completed_events[0]
+    assert event.usage is not None
+    assert isinstance(event.usage, dict)
+    assert event.usage.get("prompt_tokens", 0) > 0
+    assert event.usage.get("completion_tokens", 0) > 0
+    assert event.usage.get("total_tokens", 0) > 0
+
+
 @pytest.mark.vcr()
 def test_llm_emits_call_failed_event():
     received_events = []

lib/devtools/src/crewai_devtools/__init__.py

@@ -1,3 +1,3 @@
 """CrewAI development tools."""
 
-__version__ = "1.12.2"
+__version__ = "1.13.0a5"

lib/devtools/src/crewai_devtools/cli.py

@@ -156,6 +156,33 @@ def update_version_in_file(file_path: Path, new_version: str) -> bool:
     return False
 
 
+def update_pyproject_version(file_path: Path, new_version: str) -> bool:
+    """Update the [project] version field in a pyproject.toml file.
+
+    Args:
+        file_path: Path to pyproject.toml file.
+        new_version: New version string.
+
+    Returns:
+        True if version was updated, False otherwise.
+    """
+    if not file_path.exists():
+        return False
+
+    content = file_path.read_text()
+    new_content = re.sub(
+        r'^(version\s*=\s*")[^"]+(")',
+        rf"\g<1>{new_version}\2",
+        content,
+        count=1,
+        flags=re.MULTILINE,
+    )
+    if new_content != content:
+        file_path.write_text(new_content)
+        return True
+    return False
+
+
 _DEFAULT_WORKSPACE_PACKAGES: Final[list[str]] = [
     "crewai",
     "crewai-tools",
@@ -1045,10 +1072,84 @@ def _update_enterprise_crewai_dep(pyproject_path: Path, version: str) -> bool:
     return False
 
 
+_DEPLOYMENT_TEST_REPO: Final[str] = "crewAIInc/crew_deployment_test"
+
 _PYPI_POLL_INTERVAL: Final[int] = 15
 _PYPI_POLL_TIMEOUT: Final[int] = 600
 
 
+def _update_deployment_test_repo(version: str, is_prerelease: bool) -> None:
+    """Update the deployment test repo to pin the new crewai version.
+
+    Clones the repo, updates the crewai[tools] pin in pyproject.toml,
+    regenerates the lockfile, commits, and pushes directly to main.
+
+    Args:
+        version: New crewai version string.
+        is_prerelease: Whether this is a pre-release version.
+    """
+    console.print(
+        f"\n[bold cyan]Updating {_DEPLOYMENT_TEST_REPO} to {version}[/bold cyan]"
+    )
+
+    with tempfile.TemporaryDirectory() as tmp:
+        repo_dir = Path(tmp) / "crew_deployment_test"
+        run_command(["gh", "repo", "clone", _DEPLOYMENT_TEST_REPO, str(repo_dir)])
+        console.print(f"[green]✓[/green] Cloned {_DEPLOYMENT_TEST_REPO}")
+
+        pyproject = repo_dir / "pyproject.toml"
+        content = pyproject.read_text()
+        new_content = re.sub(
+            r'"crewai\[tools\]==[^"]+"',
+            f'"crewai[tools]=={version}"',
+            content,
+        )
+        if new_content == content:
+            console.print(
+                "[yellow]Warning:[/yellow] No crewai[tools] pin found to update"
+            )
+            return
+        pyproject.write_text(new_content)
+        console.print(f"[green]✓[/green] Updated crewai[tools] pin to {version}")
+
+        lock_cmd = [
+            "uv",
+            "lock",
+            "--refresh-package",
+            "crewai",
+            "--refresh-package",
+            "crewai-tools",
+        ]
+        if is_prerelease:
+            lock_cmd.append("--prerelease=allow")
+
+        max_retries = 10
+        for attempt in range(1, max_retries + 1):
+            try:
+                run_command(lock_cmd, cwd=repo_dir)
+                break
+            except subprocess.CalledProcessError:
+                if attempt == max_retries:
+                    console.print(
+                        f"[red]Error:[/red] uv lock failed after {max_retries} attempts"
+                    )
+                    raise
+                console.print(
+                    f"[yellow]uv lock failed (attempt {attempt}/{max_retries}),"
+                    f" retrying in {_PYPI_POLL_INTERVAL}s...[/yellow]"
+                )
+                time.sleep(_PYPI_POLL_INTERVAL)
+        console.print("[green]✓[/green] Lockfile updated")
+
+        run_command(["git", "add", "pyproject.toml", "uv.lock"], cwd=repo_dir)
+        run_command(
+            ["git", "commit", "-m", f"chore: bump crewai to {version}"],
+            cwd=repo_dir,
+        )
+        run_command(["git", "push"], cwd=repo_dir)
+        console.print(f"[green]✓[/green] Pushed to {_DEPLOYMENT_TEST_REPO}")
+
+
 def _wait_for_pypi(package: str, version: str) -> None:
     """Poll PyPI until a specific package version is available.
 
@@ -1141,6 +1242,11 @@ def _release_enterprise(version: str, is_prerelease: bool, dry_run: bool) -> Non
 
             pyproject = pkg_dir / "pyproject.toml"
             if pyproject.exists():
+                if update_pyproject_version(pyproject, version):
+                    console.print(
+                        f"[green]✓[/green] Updated version in: "
+                        f"{pyproject.relative_to(repo_dir)}"
+                    )
                 if update_pyproject_dependencies(
                     pyproject, version, extra_packages=list(_ENTERPRISE_EXTRA_PACKAGES)
                 ):
@@ -1159,7 +1265,35 @@ def _release_enterprise(version: str, is_prerelease: bool, dry_run: bool) -> Non
         _wait_for_pypi("crewai", version)
 
         console.print("\nSyncing workspace...")
-        run_command(["uv", "sync"], cwd=repo_dir)
+        sync_cmd = [
+            "uv",
+            "sync",
+            "--refresh-package",
+            "crewai",
+            "--refresh-package",
+            "crewai-tools",
+            "--refresh-package",
+            "crewai-files",
+        ]
+        if is_prerelease:
+            sync_cmd.append("--prerelease=allow")
+
+        max_retries = 10
+        for attempt in range(1, max_retries + 1):
+            try:
+                run_command(sync_cmd, cwd=repo_dir)
+                break
+            except subprocess.CalledProcessError:
+                if attempt == max_retries:
+                    console.print(
+                        f"[red]Error:[/red] uv sync failed after {max_retries} attempts"
+                    )
+                    raise
+                console.print(
+                    f"[yellow]uv sync failed (attempt {attempt}/{max_retries}),"
+                    f" retrying in {_PYPI_POLL_INTERVAL}s...[/yellow]"
+                )
+                time.sleep(_PYPI_POLL_INTERVAL)
         console.print("[green]✓[/green] Workspace synced")
 
         # --- branch, commit, push, PR ---
@@ -1175,7 +1309,7 @@ def _release_enterprise(version: str, is_prerelease: bool, dry_run: bool) -> Non
         run_command(["git", "push", "-u", "origin", branch_name], cwd=repo_dir)
         console.print("[green]✓[/green] Branch pushed")
 
-        run_command(
+        pr_url = run_command(
             [
                 "gh",
                 "pr",
@@ -1192,6 +1326,7 @@ def _release_enterprise(version: str, is_prerelease: bool, dry_run: bool) -> Non
             cwd=repo_dir,
         )
         console.print("[green]✓[/green] Enterprise bump PR created")
+        console.print(f"[cyan]PR URL:[/cyan] {pr_url}")
 
         _poll_pr_until_merged(branch_name, "enterprise bump PR", repo=enterprise_repo)
 
@@ -1558,7 +1693,18 @@ def tag(dry_run: bool, no_edit: bool) -> None:
     is_flag=True,
     help="Skip the enterprise release phase",
 )
-def release(version: str, dry_run: bool, no_edit: bool, skip_enterprise: bool) -> None:
+@click.option(
+    "--skip-to-enterprise",
+    is_flag=True,
+    help="Skip phases 1 & 2, run only the enterprise release phase",
+)
+def release(
+    version: str,
+    dry_run: bool,
+    no_edit: bool,
+    skip_enterprise: bool,
+    skip_to_enterprise: bool,
+) -> None:
     """Full release: bump versions, tag, and publish a GitHub release.
 
     Combines bump and tag into a single workflow. Creates a version bump PR,
@@ -1571,11 +1717,19 @@ def release(version: str, dry_run: bool, no_edit: bool, skip_enterprise: bool) -
         dry_run: Show what would be done without making changes.
         no_edit: Skip editing release notes.
         skip_enterprise: Skip the enterprise release phase.
+        skip_to_enterprise: Skip phases 1 & 2, run only the enterprise release phase.
     """
     try:
         check_gh_installed()
 
-        if not skip_enterprise:
+        if skip_enterprise and skip_to_enterprise:
+            console.print(
+                "[red]Error:[/red] Cannot use both --skip-enterprise "
+                "and --skip-to-enterprise"
+            )
+            sys.exit(1)
+
+        if not skip_enterprise or skip_to_enterprise:
             missing: list[str] = []
             if not _ENTERPRISE_REPO:
                 missing.append("ENTERPRISE_REPO")
@@ -1594,6 +1748,15 @@ def release(version: str, dry_run: bool, no_edit: bool, skip_enterprise: bool) -
         cwd = Path.cwd()
         lib_dir = cwd / "lib"
 
+        is_prerelease = _is_prerelease(version)
+
+        if skip_to_enterprise:
+            _release_enterprise(version, is_prerelease, dry_run)
+            console.print(
+                f"\n[green]✓[/green] Enterprise release [bold]{version}[/bold] complete!"
+            )
+            return
+
         if not dry_run:
             console.print("Checking git status...")
             check_git_clean()
@@ -1687,7 +1850,8 @@ def release(version: str, dry_run: bool, no_edit: bool, skip_enterprise: bool) -
 
         if not dry_run:
             _create_tag_and_release(tag_name, release_notes, is_prerelease)
-            _trigger_pypi_publish(tag_name, wait=not skip_enterprise)
+            _trigger_pypi_publish(tag_name, wait=True)
+            _update_deployment_test_repo(version, is_prerelease)
 
         if not skip_enterprise:
             _release_enterprise(version, is_prerelease, dry_run)

uv.lock

@@ -1243,7 +1243,7 @@ requires-dist = [
     { name = "json-repair", specifier = "~=0.25.2" },
     { name = "json5", specifier = "~=0.10.0" },
     { name = "jsonref", specifier = "~=1.1.0" },
-    { name = "lancedb", specifier = ">=0.29.2" },
+    { name = "lancedb", specifier = ">=0.29.2,<0.30.1" },
     { name = "litellm", marker = "extra == 'litellm'", specifier = ">=1.74.9,<=1.82.6" },
     { name = "mcp", specifier = "~=1.26.0" },
     { name = "mem0ai", marker = "extra == 'mem0'", specifier = "~=0.1.94" },
@@ -4275,7 +4275,7 @@ wheels = [
 
 [[package]]
 name = "nltk"
-version = "3.9.3"
+version = "3.9.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "click" },
@@ -4283,9 +4283,9 @@ dependencies = [
     { name = "regex" },
     { name = "tqdm" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/e1/8f/915e1c12df07c70ed779d18ab83d065718a926e70d3ea33eb0cd66ffb7c0/nltk-3.9.3.tar.gz", hash = "sha256:cb5945d6424a98d694c2b9a0264519fab4363711065a46aa0ae7a2195b92e71f", size = 2923673, upload-time = "2026-02-24T12:05:53.833Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/74/a1/b3b4adf15585a5bc4c357adde150c01ebeeb642173ded4d871e89468767c/nltk-3.9.4.tar.gz", hash = "sha256:ed03bc098a40481310320808b2db712d95d13ca65b27372f8a403949c8b523d0", size = 2946864, upload-time = "2026-03-24T06:13:40.641Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c2/7e/9af5a710a1236e4772de8dfcc6af942a561327bb9f42b5b4a24d0cf100fd/nltk-3.9.3-py3-none-any.whl", hash = "sha256:60b3db6e9995b3dd976b1f0fa7dec22069b2677e759c28eb69b62ddd44870522", size = 1525385, upload-time = "2026-02-24T12:05:46.54Z" },
+    { url = "https://files.pythonhosted.org/packages/9d/91/04e965f8e717ba0ab4bdca5c112deeab11c9e750d94c4d4602f050295d39/nltk-3.9.4-py3-none-any.whl", hash = "sha256:f2fa301c3a12718ce4a0e9305c5675299da5ad9e26068218b69d692fda84828f", size = 1552087, upload-time = "2026-03-24T06:13:38.47Z" },
 ]
 
 [[package]]
@@ -6235,14 +6235,14 @@ wheels = [
 
 [[package]]
 name = "pypdf"
-version = "6.9.1"
+version = "6.9.2"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "typing-extensions", marker = "python_full_version < '3.11'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/f9/fb/dc2e8cb006e80b0020ed20d8649106fe4274e82d8e756ad3e24ade19c0df/pypdf-6.9.1.tar.gz", hash = "sha256:ae052407d33d34de0c86c5c729be6d51010bf36e03035a8f23ab449bca52377d", size = 5311551, upload-time = "2026-03-17T10:46:07.876Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/31/83/691bdb309306232362503083cb15777491045dd54f45393a317dc7d8082f/pypdf-6.9.2.tar.gz", hash = "sha256:7f850faf2b0d4ab936582c05da32c52214c2b089d61a316627b5bfb5b0dab46c", size = 5311837, upload-time = "2026-03-23T14:53:27.983Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/f9/f4/75543fa802b86e72f87e9395440fe1a89a6d149887e3e55745715c3352ac/pypdf-6.9.1-py3-none-any.whl", hash = "sha256:f35a6a022348fae47e092a908339a8f3dc993510c026bb39a96718fc7185e89f", size = 333661, upload-time = "2026-03-17T10:46:06.286Z" },
+    { url = "https://files.pythonhosted.org/packages/a5/7e/c85f41243086a8fe5d1baeba527cb26a1918158a565932b41e0f7c0b32e9/pypdf-6.9.2-py3-none-any.whl", hash = "sha256:662cf29bcb419a36a1365232449624ab40b7c2d0cfc28e54f42eeecd1fd7e844", size = 333744, upload-time = "2026-03-23T14:53:26.573Z" },
 ]
 
 [[package]]