Oxygen AI Positron Enterprise（英語版） リリース

Oxygen AI Positron for Desktop

• Intercept AI Events via Hooks - AI Positron now includes a hook system for intercepting AI events to help monitor, govern, and shape AI behavior, giving you greater visibility and control over AI-driven workflows. This lets you audit AI actions, enforce organizational policies, collect usage metrics, and block risky operations before they execute.
• Configure Scripts for AI Hook Events - A new Oxygen AI Positron Hooks add-on lets you configure external scripts to run when specific AI hook events occur. Using a JSON configuration file, you can connect AI lifecycle events to your own tooling for auditing, policy enforcement, analytics, and workflow automation.
• Automatic Conversation Compaction - A new automatic conversation compaction feature allows the AI Positron to summarize older portions of long conversations (when the allowed context usage reaches 70%) and replace it with a concise recap, allowing your conversation with the AI to continue without losing context. The automatic system can be disabled via an option in the Preferences page. You can also manually trigger a compaction by using the Context Consumption button in the footer of the AI Positron view.
• Dev Agent Chat Mode - You can use the new Dev Agent chat mode to chat with the AI about your XML development project, set goals, or give instructions. The AI can help with XSL/XSLT, Schematron, DTDs, XML Schemas, XPath, validation, and transformations. It explores your project to complete even the most complex tasks efficiently, proposes a plan, and can make project-wide changes on your behalf, while you decide what to keep.
• New Development Skills - New specialized skills are available to enhance AI assistance for various development vocabularies, including XSLT, XSD, Schematron, and XSpec. These skills help the AI generate more accurate, error-free code suggestions, making development faster and more efficient.
• New OpenAI Responses Enterprise Connector - A new OpenAI Responses enterprise connector is now available. It uses the OpenAI Responses API, which is required for newer reasoning models such as GPT-5.5 Thinking.
• Reasoning Displayed in Chat for OpenAI Thinking Models - When you use OpenAI thinking models (such as GPT-5.5 Thinking) through the Positron Service or an OpenAI Responses enterprise connector, the model's reasoning is now displayed in a collapsible section in the chat.
• Connect to MCP Servers that Use OAuth Authentication - Connections to Model Context Protocol (MCP) HTTP servers that use OAuth are now supported. You can also see a summary of the servers that are running or have failed in the Connection status section and it includes log in/out buttons.
• Load Skills from Zip Archives - You can now pack skill folders as zip archives (with the .skill file extension) to distribute and share custom skills as single files. Once placed in the configured skill folders, these packages are automatically recognized and loaded by AI Positron.
• Fix Multiple Validation Problems at Once - You can now use the AI Positron Fix action on multiple selected validation problems in the Results pane, so you can resolve them in a single operation.
• AWS Bedrock: Connect Using AWS CLI Credentials (Profile/SSO) - The AWS Bedrock connector now supports AWS Credentials (Profile/SSO) as a more secure method than API key authentication. This option requires AWS CLI v2 to be installed on your machine and supports SSO and any other credential type that the AWS CLI supports.
• New AI Models Available - You can now choose from more AI models for AI-assisted tasks:
  • GPT-5.4 Thinking, GPT-5.5, and GPT-5.5 Thinking through the Positron Service or an OpenAI Responses enterprise connection.
  • Claude Opus 4.8 Thinking and Claude Opus 4.8, well-suited for agentic workflows and complex multi-step tasks.
  • Gemini 3.5 Flash through the Positron Service or the VertexAI enterprise connection.
• Manage Available Models for Configured AI Connections - Administrators can now control which models are available to users for configured AI connections, including OpenAI, Claude, and Gemini connectors. You can add or remove models from the connection-specific list, define a recommended default model, and include additional models that are not part of the predefined list.
• Editor Variables Supported in Prompt Reference When Creating Custom AI Actions - When creating custom AI actions, you can now use Oxygen editor variables when defining the prompt_ref property (for example, to reference content from the user's home folder) and the content will be expanded automatically.
• New AI Tool Function for Evaluating XPath Expressions - AI Positron now provides the evaluate_xpath tool function to the LLM, enabling it to retrieve nodes against a document resource using XPath 2.0 expressions. The result includes the matched element and its location within the document, giving you a faster and simpler way to access structured content.
• New XPath AI Functions - ai:chat() and ai:resolve-action() - Two new AI extension functions can be called from XPath, and any XPath-enabled language (XSLT, XQuery, XProc, XSpec, Schematron, Schematron Quick Fix, etc.). The ai:chat function runs a full chat completion. It takes a conversation (and optional tool definitions and model settings) and returns the complete message history, allowing multi-turn exchanges to be continued. The ai:resolve-action function turns a saved AI Positron action into the messages and options that are passed to the ai:chat function, letting you separate an action's definition from its execution. Like the other AI functions, they can be enabled or disabled from the XPath Functions preferences page.

Oxygen AI Positron for Eclipse

• OpenAI Responses Connector for Reasoning Models - You can now configure an OpenAI Responses enterprise connector to use the OpenAI Responses API. This connector is required for newer OpenAI reasoning models (such as GPT-5.5 Thinking) and keeps requests from being stored on OpenAI's servers.
• New Development Skills - New specialized skills are available to enhance AI assistance for various development vocabularies, including XSLT, XSD, Schematron, and XSpec. These skills help the AI generate more accurate, error-free code suggestions, making development faster and more efficient.
• Load Skills from Zip Archives - You can now pack skill folders as zip archives (with the .skill file extension) to distribute and share custom skills as single files. Once placed in the configured skill folders, these packages are automatically recognized and loaded by AI Positron.
• New AI Models Available - You can now choose from more AI models for AI-assisted tasks:
  • GPT-5.4 Thinking, GPT-5.5, and GPT-5.5 Thinking through the OpenAI Responses connector.
  • Claude Opus 4.8 Thinking and Claude Opus 4.8, well-suited for agentic workflows and complex multi-step tasks.
  • Gemini 3.5 Flash through the Gemini connector.
• Editor Variables Supported in Prompt Reference When Creating Custom AI Actions - When creating custom AI actions, you can now use Oxygen editor variables when defining the prompt_ref property (for example, to reference content from the user's home folder) and the content will be expanded automatically.
• New AI Tool Function for Evaluating XPath Expressions - AI Positron now provides the evaluate_xpath tool function to the LLM, enabling it to retrieve nodes against a document resource using XPath 2.0 expressions. The result includes the matched element and its location within the document, giving you a faster and simpler way to access structured content.

Oxygen AI Positron for Content Fusion

• Intercept AI Events via Hooks - AI Positron now includes a hook system for intercepting AI events to help monitor, govern, and shape AI behavior, giving you greater visibility and control over AI-driven workflows. This lets you audit AI actions, enforce organizational policies, collect usage metrics, and block risky operations before they execute.
• Automatic Conversation Compaction - A new automatic conversation compaction feature allows the AI Positron to summarize older portions of long conversations (when the allowed context usage reaches 70%) and replace it with a concise recap, allowing your conversation with the AI to continue without losing context. The automatic system can be disabled via an option in the Administration page. You can also manually trigger a compaction by using the Context Consumption button in the footer of the AI Positron view.
• AI Automatically Saves Custom Actions, Skills, and Prompts in the Current Project - When you ask the AI to create a custom AI action, skill, or context prompt, it now automatically determines the correct location to save it within your current project. This removes the need to manually specify or organize where these customizations are stored.
• New OpenAI Responses Enterprise Connector - A new OpenAI Responses enterprise connector is now available. It uses the OpenAI Responses API, which is required for newer reasoning models such as GPT-5.5 Thinking.
• Reasoning Displayed in Chat for OpenAI Thinking Models - When you use OpenAI thinking models (such as GPT-5.5 Thinking) through the Positron Service or an OpenAI Responses enterprise connector, the model's reasoning is now displayed in a collapsible section in the chat.
• Load Skills from Zip Archives - You can reference an archive that contains custom AI actions, agents, and skills in the Web Author admin pages. Or you can place the custom AI actions, agents and skills in a special folder located in the Content Fusion project (.contentfusion/positron-ai-actions). These skill folders can now be packed as zip archives (with the .skill file extension) and the AI Positron will dynamically load the skills from the archives.
• New AI Models Available - You can now choose from more AI models for AI-assisted tasks:
  • GPT-5.4 Thinking, GPT-5.5, and GPT-5.5 Thinking through the OpenAI Responses enterprise connection.
  • Claude Opus 4.8 Thinking and Claude Opus 4.8, well-suited for agentic workflows and complex multi-step tasks.
  • Gemini 3.5 Flash through the VertexAI enterprise connection.
• Manage Available Models for Configured AI Connections - Administrators can now control which models are available to users for configured AI connections, including OpenAI, Claude, and Gemini connectors. You can add or remove models from the connection-specific list, define a recommended default model, and include additional models that are not part of the predefined list.

Oxygen AI Positron for Web Author

• Intercept AI Events via Hooks - AI Positron now includes a hook system for intercepting AI events to help monitor, govern, and shape AI behavior, giving you greater visibility and control over AI-driven workflows. This lets you audit AI actions, enforce organizational policies, collect usage metrics, and block risky operations before they execute.
• Automatic Conversation Compaction - A new automatic conversation compaction feature allows the AI Positron to summarize older portions of long conversations (when the allowed context usage reaches 70%) and replace it with a concise recap, allowing your conversation with the AI to continue without losing context. The automatic system can be disabled via an option in the Administration page. You can also manually trigger a compaction by using the Context Consumption button in the footer of the AI Positron view.
• New OpenAI Responses Enterprise Connector - A new OpenAI Responses enterprise connector is now available. It uses the OpenAI Responses API, which is required for newer reasoning models such as GPT-5.5 Thinking.
• Reasoning Displayed in Chat for OpenAI Thinking Models - When you use OpenAI thinking models (such as GPT-5.5 Thinking) through the Positron Service or an OpenAI Responses enterprise connector, the model's reasoning is now displayed in a collapsible section in the chat.
• Load Skills from Zip Archives - You can reference an archive that contains custom AI actions, agents, and skills in the Web Author admin pages. These skill folders can now be packed as zip archives (with the .skill file extension) and the AI Positron will dynamically load the skills from the archives.
• New AI Models Available - You can now choose from more AI models for AI-assisted tasks:
  • GPT-5.4 Thinking, GPT-5.5, and GPT-5.5 Thinking through the OpenAI Responses enterprise connection.
  • Claude Opus 4.8 Thinking and Claude Opus 4.8, well-suited for agentic workflows and complex multi-step tasks.
  • Gemini 3.5 Flash through the VertexAI enterprise connection.
• Manage Available Models for Configured AI Connections - Administrators can now control which models are available to users for configured AI connections, including OpenAI, Claude, and Gemini connectors. You can add or remove models from the connection-specific list, define a recommended default model, and include custom models.

Oxygen AI Positron for Desktop

• Extend AI Capabilities Using Skills - Added the ability to define and use skills (based on the Anthropic Agent Skills) to extend the AI agent's capabilities by creating reusable, specialized modules for specific tasks. You can use them for automating repetitive workflows, integrating with external tools and APIs, standardizing best practices, and enabling the assistant to handle domain-specific or proprietary processes. From simple utilities to complex integrations, skills let you customize the AI agent to match your unique business needs and accelerate productivity.
• Built-in AI Skills - Several built-in skills are now available to help when you want the LLM to:
  • Add review comments to your document.
  • Translate XML content into another language while preserving markup and technical elements.
  • Apply structured persona-based analysis to evaluate a task through multiple audience perspectives.
  • Get help with AI Positron features by automatically reading the user guide documentation.
• Claude Models with Adaptive Thinking - Two new Claude models are available: Claude Sonnet 4.6 Thinking and Claude Opus 4.6 Thinking. These models use adaptive thinking to automatically determine how much reasoning is needed for each request, spending more time thinking through complex problems while responding quickly to simpler queries. When using these models, the chat interface displays a collapsible summary of the model's thinking process in real time.
• Personas Chat Mode - You can now use the Personas AI chat mode to review a task from multiple audience perspectives and get a more balanced recommendation. If you do not specify personas, the agent automatically defines 2-4 relevant personas, analyzes the content from their perspectives, identifies conflicting viewpoints, and tries to reach a compromise before providing a final recommendation.
• Smarter AI Awareness of Changes in Your Working Context - When you continue a chat after changing the selected content or switching to a different file, the AI can now follow those context changes better and respond based on what you are currently working on. This helps reduce confusion in ongoing conversations and makes follow-up requests feel more accurate and natural.
• Binary Resource Insights - You can now ask the AI to analyze binary resources such as images, PDF, Word, Excel, and PowerPoint files without first attaching them in the chat. This makes it easier to generate content based on files in a folder or validate whether referenced images fit the surrounding context.
• Context Window Size Indicator - A new indicator in the AI Positron footer gives you visibility into the model's context window size so you can understand how much context is available for your conversation. The indicator includes a breakdown of the total context window and how much of that context is currently in use.
• Attach Excel Documents - You can now attach Excel (XLSX) documents in the AI Positron chat view to provide additional context for your conversation with the AI.
• Analyze Referenced SVG Images - When using AI actions such as "Generate image alternate text" or interacting with the AI in the chat, the LLM is now given access to the referenced SVG images (in addition to JPEG and PNG images). This allows the AI to provide feedback based on the SVG image content.
• Review Changes Support for Copy/Move/Rename AI Operations - The Review Changes sandbox now tracks copy, move, and rename operations on files and folders performed by the AI. Each affected resource appears as a single entry in the Waiting for Review panel (even if the file was both modified and moved or renamed during the conversation), so you can clearly see what happened and accept or reject the changes. Rejected operations restore files and folders to their original state.
• Configure Reasoning Effort Level for Microsoft Azure Connections - For Microsoft Azure connections, you can now choose the default reasoning effort level (low, medium, or high) to use when executing actions or chat requests via the new Reasoning Effort parameter.
• Support for Claude Deployments in Microsoft Azure Connections - Claude (Anthropic) models hosted on Azure AI Foundry can now be used directly from your Microsoft Azure AI connections.
• New Gemini Models Available for Oxygen AI Positron Service - The Gemini 3 Flash, 3.1 Pro, and 3.1 Flash Lite models are now available in the list of models to choose from when using the Oxygen AI Positron Service.
• Define Multiple AI Connections in JSON Configuration Files - If you configure your AI services using a JSON configuration file, you can now define an array of multiple connections. This provides your users with the flexibility to choose their preferred AI service connection directly from the interface.
• Improved Tools Approval - When prompted to run potentially unsafe tools (for example, tools provided by Model Context Protocol integrations), you can now choose to always approve specific tools. The confirmation prompt no longer times out, preventing automatic cancellations. Additionally, the entire UI of the confirmation request has been redesigned to be more intuitive and easier to read.
• Model Context Protocol (MCP) Improvements - The Model Context Protocol (MCP) SDK has been updated to version 1.1.1. You can now use Oxygen editor variables for all parameter values in an AI Positron MCP configuration file to easily create dynamic and portable configurations. Additionally, you can specify the cwd (current working directory) property for local MCP connections to ensure your commands execute in the correct local context.
• Ask for Website Access - You are now asked directly in the chat view for permission if the LLM attempts to read content from any website. You can choose to always allow connections to a certain website host.

Oxygen AI Positron for Eclipse

• Extend AI Capabilities Using Skills - Added the ability to define and use skills (based on the Anthropic Agent Skills) to extend the AI agent's capabilities by creating reusable, specialized modules for specific tasks. You can use them for automating repetitive workflows, integrating with external tools and APIs, standardizing best practices, and enabling the assistant to handle domain-specific or proprietary processes. From simple utilities to complex integrations, skills let you customize the AI agent to match your unique business needs and accelerate productivity.
• Built-in AI Skills - Several built-in skills are now available to help when you want the LLM to:
  • Add review comments to your document.
  • Translate XML content into another language while preserving markup and technical elements.
  • Apply structured persona-based analysis to evaluate a task through multiple audience perspectives.
  • Get help with AI Positron features by automatically reading the user guide documentation.
• Claude Models with Adaptive Thinking - Two new Claude models are available: Claude Sonnet 4.6 Thinking and Claude Opus 4.6 Thinking. These models use adaptive thinking to automatically determine how much reasoning is needed for each request, spending more time thinking through complex problems while responding quickly to simpler queries.
• Smarter AI Awareness of Changes in Your Working Context - When you continue a chat after changing the selected content or switching to a different file, the AI can now follow those context changes better and respond based on what you are currently working on. This helps reduce confusion in ongoing conversations and makes follow-up requests feel more accurate and natural.
• Binary Resource Insights - You can now ask the AI to analyze binary resources such as images, PDF, Word, Excel, and PowerPoint files without first attaching them in the chat. This makes it easier to generate content based on files in a folder or validate whether referenced images fit the surrounding context.
• Attach Excel Documents - You can now attach Excel (XLSX) documents in the AI Positron chat view to provide additional context for your conversation with the AI.
• Analyze Referenced SVG Images - When using AI actions such as "Generate image alternate text" or interacting with the AI in the chat, the LLM is now given access to the referenced SVG images (in addition to JPEG and PNG images). This allows the AI to provide feedback based on the SVG image content.
• Support for Claude Deployments in Microsoft Azure Connections - Claude (Anthropic) models hosted on Azure AI Foundry can now be used directly from your Microsoft Azure AI connections.
• Configure Multiple Action, Agent, and Skill Folders - You can now configure a comma-separated list of folders that the AI Positron will use to load actions, agents, and skills. This can be done in the Actions section of the AI Positron preferences page.

Oxygen AI Positron for Content Fusion

• Select the LLM Model Directly from the Chat View - Depending on the configured LLM connection, you can now select the LLM model to use when chatting or invoking AI actions using a drop-down control in the chat box.
• Extend AI Capabilities Using Skills - Added the ability to define and use skills (based on the Anthropic Agent Skills) to extend the AI agent's capabilities by creating reusable, specialized modules for specific tasks. You can use them for automating repetitive workflows, integrating with external tools and APIs, standardizing best practices, and enabling the assistant to handle domain-specific or proprietary processes. From simple utilities to complex integrations, skills let you customize the AI agent to match your unique business needs and accelerate productivity.
• Built-in AI Skills - Several built-in skills are now available to help when you want the LLM to:
  • Add review comments to your document.
  • Translate XML content into another language while preserving markup and technical elements.
  • Apply structured persona-based analysis to evaluate a task through multiple audience perspectives.
• Load Actions, Skills, and Project Context from Content Fusion Projects - If a Content Fusion project contains an ai-positron-actions folder in the project root (or in .contentfusion/ai-positron-actions), AI Positron automatically loads custom actions and skills from it. If the folder also contains an ai-context.md file, its contents are automatically used as the project context prompt for new chats. Updates to these resources are applied when you start a new chat.
• Claude Models with Adaptive Thinking - Two new Claude models are available: Claude Sonnet 4.6 Thinking and Claude Opus 4.6 Thinking. These models use adaptive thinking to automatically determine how much reasoning is needed for each request, spending more time thinking through complex problems while responding quickly to simpler queries. When using these models, the chat interface displays a collapsible summary of the model's thinking process in real time. The thinking summary is also preserved and restored when you load conversations from history.
• Personas Chat Mode - You can now use the Personas AI chat mode to review a task from multiple audience perspectives and get a more balanced recommendation. If you do not specify personas, the agent automatically defines 2-4 relevant personas, analyzes the content from their perspectives, identifies conflicting viewpoints, and tries to reach a compromise before providing a final recommendation.
• Smarter AI Awareness of Changes in Your Working Context - When you continue a chat after changing the selected content or switching to a different file, the AI can now follow those context changes better and respond based on what you are currently working on. This helps reduce confusion in ongoing conversations and makes follow-up requests feel more accurate and natural.
• Binary Resource Insights - You can now ask the AI to analyze binary resources such as images, PDF, Word, Excel, and PowerPoint files without first attaching them in the chat. This makes it easier to generate content based on files in a folder or validate whether referenced images fit the surrounding context.
• Context Window Size Indicator - A new indicator in the AI Positron footer gives you visibility into the model's context window size so you can understand how much context is available for your conversation. The indicator includes a breakdown of the total context window and how much of that context is currently in use.
• Easily Review AI-Modified Files - Files created or modified by the AI now appear directly in the chat area, where you can review the changes. You can compare the original content with the AI-updated version or directly open a changed file to make corrections.
• Attach Excel Documents - You can now attach Excel (XLSX) documents in the AI Positron chat view to provide additional context for your conversation with the AI.
• Analyze Referenced SVG Images - When using AI actions such as "Generate image alternate text" or interacting with the AI in the chat, the LLM is now given access to the referenced SVG images (in addition to JPEG and PNG images). This allows the AI to provide feedback based on the SVG image content.
• Image References Expanded in Selected Content - When you select content that includes image references in the visual editor, the referenced images are included along with the text. This allows you to ask the AI agent to review the text and images together, for example, to check whether the image and its accompanying text work well together.
• Configure Reasoning Effort Level for Microsoft Azure Connections - For Microsoft Azure connections, you can now choose the default reasoning effort level (low, medium, or high) to use when executing actions or chat requests via the new Reasoning Effort parameter.
• Support for Claude Deployments in Microsoft Azure Connections - Claude (Anthropic) models hosted on Azure AI Foundry can now be used directly from your Microsoft Azure AI connections.
• Edit Markdown Documents While Vibe Writing - When vibe writing, AI Positron can now retrieve content from and modify the current document, whether it is an XML or Markdown document.

Oxygen AI Positron for Web Author

• Select LLM Model in Chat View - Depending on the configured LLM connection, you can now select the LLM model to use when chatting or invoking AI actions using a drop-down control in the chat box.
• Extend AI Capabilities Using Skills - Added the ability to define and use skills (based on the Anthropic Agent Skills) to extend the AI agent's capabilities by creating reusable, specialized modules for specific tasks. You can use them for automating repetitive workflows, integrating with external tools and APIs, standardizing best practices, and enabling the assistant to handle domain-specific or proprietary processes. From simple utilities to complex integrations, skills let you customize the AI agent to match your unique business needs and accelerate productivity.
• Built-in AI Skills - Several built-in skills are now available to help when you want the LLM to:
  • Add review comments to your document.
  • Translate XML content into another language while preserving markup and technical elements.
  • Apply structured persona-based analysis to evaluate a task through multiple audience perspectives.
• Claude Models with Adaptive Thinking - Two new Claude models are available: Claude Sonnet 4.6 Thinking and Claude Opus 4.6 Thinking. These models use adaptive thinking to automatically determine how much reasoning is needed for each request, spending more time thinking through complex problems while responding quickly to simpler queries. When using these models, the chat interface displays a collapsible summary of the model's thinking process in real time. The thinking summary is also preserved and restored when you load conversations from history.
• Personas Chat Mode - You can now use the Personas AI chat mode to review a task from multiple audience perspectives and get a more balanced recommendation. If you do not specify personas, the agent automatically defines 2-4 relevant personas, analyzes the content from their perspectives, identifies conflicting viewpoints, and tries to reach a compromise before providing a final recommendation.
• Smarter AI Awareness of Changes in Your Working Context - When you continue a chat after changing the selected content or switching to a different file, the AI can now follow those context changes better and respond based on what you are currently working on. This helps reduce confusion in ongoing conversations and makes follow-up requests feel more accurate and natural.
• Binary Resource Insights - You can now ask the AI to analyze binary resources such as images, PDF, Word, Excel, and PowerPoint files without first attaching them in the chat. This makes it easier to generate content based on files in a folder or validate whether referenced images fit the surrounding context.
• Context Window Size Indicator - A new indicator in the AI Positron footer gives you visibility into the model's context window size so you can understand how much context is available for your conversation. The indicator includes a breakdown of the total context window and how much of that context is currently in use.
• Easily Review AI-Modified Files - Files created or modified by the AI now appear directly in the chat area, where you can review the changes. You can compare the original content with the AI-updated version or directly open a changed file to make corrections.
• Attach Excel Documents - You can now attach Excel (XLSX) documents in the AI Positron chat view to provide additional context for your conversation with the AI.
• Analyze Referenced SVG Images - When using AI actions such as "Generate image alternate text" or interacting with the AI in the chat, the LLM is now given access to the referenced SVG images (in addition to JPEG and PNG images). This allows the AI to provide feedback based on the SVG image content.
• Image References Expanded in Selected Content - When you select content that includes image references in the visual editor, the referenced images are included along with the text. This allows you to ask the AI agent to review the text and images together, for example, to check whether the image and its accompanying text work well together.
• Configure Reasoning Effort Level for Microsoft Azure Connections - For Microsoft Azure connections, you can now choose the default reasoning effort level (low, medium, or high) to use when executing actions or chat requests via the new Reasoning Effort parameter.
• Support for Claude Deployments in Microsoft Azure Connections - Claude (Anthropic) models hosted on Azure AI Foundry can now be used directly from your Microsoft Azure AI connections.
• Edit Markdown Documents While Vibe Writing - When vibe writing, AI Positron can now retrieve content from and modify the current document, whether it is an XML or Markdown document.

Oxygen AI Positron for Desktop

• OpenAI GPT-5.2 Models Available - The flagship GPT-5.2 and GPT-5.2 Thinking models, OpenAI's flagship models for agentic tasks, are now available in the list of OpenAI models to choose from. The Thinking variant is designed for complex tasks and multi-step decisions where accuracy matters more than speed.
• Sensitive Data Masking for AI Requests - Personal or private information (such as email addresses, phone numbers, IDs, or API keys) is now automatically identified and masked in AI requests to help protect sensitive data. You can extend this protection by adding custom patterns to cover additional data types. The masked information is automatically restored in the AI response, ensuring that users see the correct content and that any generated or updated files remain accurate.
• Improved Review Experience for AI-Modified Files - All files created or modified by AI now appear directly in the chat area, so you can review the proposed changes and accept or discard them without leaving the conversation. You can also open each file individually and accept only a subset of the proposed changes while rejecting the rest.
• Optimized Token Consumption for AI Prompts and Tools - The prompts and tools that the AI uses to generate answers have been improved to consume fewer tokens/credits.
• New Planner Chat Mode for Resource Consuming Tasks - The newly-introduced Planner chat mode is available for complex, token-intensive operations (e.g. "Translate all files in the current folder" or "Find all conrefs and replace them with conkeyrefs in userguide.ditamap"). After being given the prompt, the AI explores your project, proposes a multi-step plan, and after your approval, delegates each step to the most suitable AI agent. This means large, project-wide changes can be coordinated without hitting LLM context limits, while still showing you the tool calls and results for review.
• Smarter PDF/Word Attachment Expansion in Planner Mode - When using the Planner chat mode with PDF or Word (.docx) attachments, the document is automatically split into multiple Markdown files, with embedded images extracted into a temporary folder. This allows the Planner to invoke sub-agents that can process, transform, or convert the extracted files and assets as needed (for example, converting them to other formats or generating new project structures).
• More Context Information Available in Chat - The AI chat now automatically includes additional information about your environment and editing context, such as the Oxygen application version, operating system details, current date, main file information, and caret or selection data (Oxygen 28.1 or newer).
• Tools Selector - You can now customize the AI's capabilities for each chat mode. The new Tools selector allows you to disable or enable specific tools (either added via Model Context Protocol or provided in the chat mode definition), giving you more control to fine-tune the AI's responses and behavior for your specific tasks.
• Chat Selection Includes Images - When you use the AI chat with content selected in your document, any images within that selection are now also included for the LLM to analyze along with the text.
• Create New AI Chat Modes - You can now extend the AI chat by creating your own custom chat modes. This allows you to build specialized assistants tailored to your specific needs and workflows.
• AI Positron Fix Action Turns Agentic - The AI Positron Fix action now automatically applies proposed fixes directly in your document and continues to explore additional sections when more context is needed.
• Documentation and APIs to Create New Chat Modes - You can now officially define new chat modes by creating new AI actions that contain the "output-action": "agentic" property.
• Tools to List and Invoke AI Agents - Newly added tools make it possible to create powerful new chat modes to plan complex changes and then use separate agents for each part of the plan.
• Tool to Run Command Line - A newly added tool is available in all chat modes and can be used to invoke a command line process. Note that the tool needs explicit user approval when it is invoked.
• Tool to Copy/Move/Rename Local Resources - A new tool is available in all chat modes and can be used to copy, move, or rename a local file or folder path. When used in Oxygen 28.1 and newer, if the Main Files feature is enabled in Oxygen, references to the moved or renamed resource are also automatically updated.
• Control AI Action Visibility - You can now choose whether a defined AI action is visible in the UI or only available to other AI actions as a sub-agent by using the new visibility property.
• Create AI Actions with Vision Support - You can now decide whether a custom AI action sends both content and referenced images to the LLM by using the "input-type": "markup-with-images" JSON property. This helps you include images only when they are relevant for the AI response.
• Generate Image Alternate Text Improvements - The Generate Image Alternate Text action can now be used on all images that are referenced within a document at once, while you can still use the action on individual selections as well.
• Preserve Structure and Styles for Attached Word and PDF Documents - When attaching Word (.docx) or PDF documents in the chat, their internal structure (headings, lists, tables, code blocks) and styling (bold, italic) are now preserved when being sent to the AI so that it better understands the document.
• Updated Model Context Protocol Library - The Model Context Protocol SDK that is bundled with AI Positron 8.1 has been updated to contain the latest fixes in the 0.16.0 SDK.
• Removed Support for Older Models - Support for older models (such as GPT-4, GPT-4 Turbo, GPT-4o, Sonnet 3.7, and more) has been removed. Alternatives with improved characteristics are available, such as models with higher intelligence, lower prices, or larger context windows.

Oxygen AI Positron for Eclipse

• OpenAI GPT-5.2 Models Available - The flagship GPT-5.2 and GPT-5.2 Thinking models are now available in the list of OpenAI models to choose from.
• Optimized Token Consumption for AI Prompts and Tools - The prompts and tools that the AI uses to generate answers have been improved to consume fewer tokens/credits.
• More Context Information Available in Chat - The AI chat now automatically includes additional information about your environment and editing context, such as the Oxygen application version, operating system details, current date, and caret or selection data (Oxygen 28.1 or newer).
• Control AI Action Visibility - You can now choose whether a defined AI action is visible in the UI or only available to other AI actions as a sub-agent by using the new visibility property.
• Create AI Actions with Vision Support - You can now decide whether a custom AI action sends both content and referenced images to the LLM by using the "input-type": "markup-with-images" JSON property. This helps you include images only when they are relevant for the AI response.
• Generate Image Alternate Text Improvements - The Generate Image Alternate Text action can now be used on all images that are referenced within a document at once, while you can still use the action on individual selections as well.
• Preserve Structure and Styles for Attached Word and PDF Documents - When attaching Word (.docx) or PDF documents in the chat, their internal structure (headings, lists, tables, code blocks) and styling (bold, italic) are now preserved when being sent to the AI so that it better understands the context of the document.

Oxygen AI Positron for Content Fusion

• OpenAI GPT-5.2 Models Available - The flagship GPT-5.2 and GPT-5.2 Thinking models are now available in the list of OpenAI models to choose from.
• Optimized Token Consumption for AI Prompts and Tools - The prompts and tools that the AI uses to generate answers have been improved to consume fewer tokens/credits.
• New Planner Chat Mode for Resource Consuming Tasks - The newly-introduced Planner chat mode is available for complex, token-intensive operations (e.g. "Translate all files in the current folder" or "Find all conrefs and replace them with conkeyrefs in userguide.ditamap"). After being given the prompt, the AI explores your project, proposes a multi-step plan, and, after your approval, delegates each step to the most suitable AI agent. This means large, project-wide changes can be coordinated without hitting LLM context limits, while still showing you the tool calls and results for review.
• More Context Information Available in Chat - The AI chat now automatically includes additional information about your environment and editing context, such as the Oxygen application version, current date, and caret or selection data (Content Fusion 9.0 or newer).
• Documentation and APIs to Create New Chat Modes - You can now officially define new chat modes by creating new AI actions that contain the "output-action": "agentic" property.
• Tools to List and Invoke AI Agents - Newly added tools make it possible to create powerful new chat modes to plan complex changes and then use separate agents for each part of the plan.
• List Directory Tool (Available with Content Fusion 9.1) - When using an agent chat mode, a new list_dir tool is available to list the contents of a directory, which helps the AI understand the structure of your project.
• Control AI Action Visibility - You can now choose whether a defined AI action is visible in the UI or only available to other AI actions as a sub-agent by using the new visibility property.
• Create AI Actions with Vision Support - You can now control whether a custom-defined AI action sends both content and referenced images to the LLM by using the "input-type": "markup-with-images" JSON property.
• Generate Image Alternate Text Improvements - The Generate Image Alternate Text action can now be used on all images that are referenced within a document at once, while you can still use the action on individual selections as well.
• Preserve Structure and Styles for Attached Word and PDF Documents - When attaching Word (.docx) or PDF documents in the chat, their internal structure (headings, lists, tables, code blocks) and styling (bold, italic) are now preserved when being sent to the AI so that it better understands the context of the document.

Oxygen AI Positron for Web Author

• OpenAI GPT-5.2 Models Available - The flagship GPT-5.2 and GPT-5.2 Thinking models are now available in the list of OpenAI models to choose from.
• Optimized Token Consumption for AI Prompts and Tools - prompts and tools that the AI uses to generate answers have been improved to consume fewer tokens/credits.
• New Planner Chat Mode for Resource Consuming Tasks - The newly-introduced Planner chat mode is available for complex, token-intensive operations (e.g. "Translate all files in the current folder" or "Find all conrefs and replace them with conkeyrefs in userguide.ditamap"). After being given the prompt, the AI explores your project, proposes a multi-step plan, and, after your approval, delegates each step to the most suitable AI agent. This means large, project-wide changes can be coordinated without hitting LLM context limits, while still showing you the tool calls and results for review.
• More Context Information Available in Chat - The AI chat now automatically includes additional information about your environment and editing context, such as the Oxygen application version, current date, and caret or selection data (Web Author 28.1 or newer).
• Documentation and APIs to Create New Chat Modes - You can now officially define new chat modes by creating new AI actions that contain the "output-action": "agentic" property.
• Tools to List and Invoke AI Agents - Newly added tools make it possible to create powerful new chat modes to plan complex changes and then use separate agents for each part of the plan.
• Control AI Action Visibility - You can now choose whether a defined AI action is visible in the UI or only available to other AI actions as a sub-agent by using the new visibility property.
• Create AI Actions with Vision Support - You can now control whether a custom-defined AI action sends both content and referenced images to the LLM by using the "input-type": "markup-with-images" JSON property.
• Generate Image Alternate Text Improvements - The Generate Image Alternate Text action can now be used on all images that are referenced within a document at once, while you can still use the action on individual selections as well.
• Preserve Structure and Styles for Attached Word and PDF Documents - When attaching Word (.docx) or PDF documents in the chat, their internal structure (headings, lists, tables, code blocks) and styling (bold, italic) are now preserved when being sent to the AI so that it better understands the context of the document.