Changelog

This update fixes an issue where the agent sometimes received text in an unexpected format from certain AI providers. It now ensures that text chunks are always combined correctly, leading to more reliable responses.

New examples of how people are using Hermes have been added to showcase its latest capabilities. This update brings more diverse and recent real-world applications of the agent, helping users discover fresh ideas.

A new optional skill called "grill-me" has been added. This skill helps you stress-test your ideas by asking structured questions before you start coding, ensuring your plan is solid.

The agent will no longer freeze when trying to compress a response. If compression takes too long, the agent will now send the uncompressed response and continue processing future requests. This prevents users from getting stuck waiting for a response.

We've adjusted how close buttons are displayed on tabbed panes to ensure they are always visible and clickable, even with short tab titles. This change improves usability by preventing the close button from being hidden by the tab's text.

When managing multiple skills at once, if one fails, the system now provides more detailed information about that failure. This helps the AI understand what went wrong and recover more effectively, leading to fewer attempts and less processing.

This update improves how the agent handles authentication requests. It prevents a single long-running request from blocking others, ensuring that multiple requests can be processed concurrently without interruption.

The Hy4 model from Tencent is now available for use, improving its accessibility and ensuring it works even when not connected to the live catalog. This gives users more advanced AI options for their tasks.

You can now move your Kanban boards between computers or archive them by exporting and importing them. You can also rename and delete boards directly from the desktop switcher.

The project's code formatting and style checks can now be automatically fixed, ensuring a cleaner and more consistent codebase. This makes it easier to maintain and contribute to the project by resolving minor issues automatically.

The Kanban board switcher no longer crashes when you open it. This bug, which prevented you from seeing or switching between your boards, is now fixed.

The project's code formatting and style has been automatically updated to ensure consistency and catch potential issues. This helps maintain a clean and robust codebase.

When a skill update fails because it can't find the original text, Hermes Agent now provides helpful instructions for fixing it instead of accidentally overwriting the entire skill. This makes it easier to recover from errors and ensures skills are updated correctly.

The Hermes Agent's desktop tools have been simplified and made more efficient. This update combines similar preview and project management actions, reducing the code size and making the agent faster to use.

The pet animation in the command-line interface now looks sharper and smoother in compatible terminals like Kitty and Ghostty. This makes the interactive CLI more visually appealing by using a better way to display the animated pet.

This update significantly speeds up reading data by allowing many operations to happen at the same time instead of waiting for each other. This makes the agent more responsive, especially when managing many sessions.

Pasted images in Microsoft Teams will now be correctly delivered to the agent instead of being rejected. This change ensures that images shared directly in chats are processed like file uploads, improving the user experience.

Hermes Agent will now correctly time out stalled summary streams. Previously, empty messages could prevent timeouts, leading to sessions that never ended.

The Hermes Agent now correctly picks up changes to service account files. This means that if you rotate your service account credentials, the agent will immediately start using the new, updated credentials instead of the old ones.

This update fixes issues with how prompts are remembered and how conversation progress is saved on your desktop. These changes ensure that more complex prompts are handled correctly and prevent data loss when saving chats, leading to a more reliable experience.

The on-screen heads-up display on Linux now correctly responds to clicks, including for links and the right-click menu. Links in the display will now open in your default web browser, improving usability and preventing issues with loading web content.

The project's code formatting and style checks can now be automatically fixed by running a single command. This ensures consistent code quality and makes it easier for contributors to maintain the project.

The prompt cache time for some language models has been restored to one hour. This ensures that frequently used models are cached for longer, improving their performance.

Tool-using conversations with Anthropic through LiteLLM proxies no longer crash due to a formatting error when using cached responses. This update ensures that caching works correctly for these specific tool interactions.

The project's code formatting has been automatically improved for better readability and consistency. This makes it easier for developers to understand and contribute to the project.

Code has been simplified by removing a redundant helper function for retrieving credentials. This consolidation ensures that the logic for getting credentials is handled in one place, making the code cleaner and easier to maintain.

The way Hermes Agent remembers and uses your preferences has been streamlined to be more efficient and clear. This change makes its instructions shorter and more accurate, meaning it uses less processing power with each interaction.

Files like PDFs are now displayed correctly with a download button and preview card, just like media files, ensuring a consistent experience when sharing any type of document. This fix aligns the primary file sharing method with how plain links behave, making it easier to interact with all shared files.

A download button has been added to file preview cards, making it easier to save delivered files directly to your computer. This new button works for both local and remote file sources and uses existing app features for saving and error handling.

The list of free models you can choose from now correctly updates without needing a new project release. This means unavailable models will disappear, and new ones will appear automatically.

Previewing files from remote or URL connections now works correctly. Previously, this feature silently failed for remote connections, but it has been fixed to display inline previews as intended.

The `grill-me` skill has been renamed to `plan-interrogation` and now uses a more efficient interviewing method where the agent asks multiple questions at once. This change makes the skill easier to understand and improves the interaction by presenting all available questions with recommendations based on their dependencies.

Hermes can now create interactive guides that walk you through complex, human-only setup tasks like grabbing dashboard credentials or adding secrets to your CI. This makes it easier and safer to complete those important one-time setup steps.

Hermes can now help you gather information for decisions by creating structured questionnaires. This new feature is useful when a decision requires input from someone else, allowing them to provide answers asynchronously.

The way we explain how to add interactive widgets to your desktop content has been improved. This update clarifies the process by showing you how to build them step-by-step and explains why certain style choices are automatically handled.

The `grill-me` skill has been restored to its original name, keeping a content upgrade that was part of a previous change. This ensures the skill is easily identifiable while still benefiting from improvements.

We've fixed two platform descriptions that incorrectly claimed audio would be sent as voice messages. Now, Hermes Agent will correctly convert non-voice audio formats to voice messages for Telegram and Feishu, ensuring a better user experience. Additionally, we've verified and corrected many other platform-specific text formatting rules to be more accurate.

This update automatically fixes code formatting and linting issues in the JavaScript parts of Hermes Agent. This helps keep the code clean and consistent, making it easier to maintain.

The way Hermes Agent explains how to send files has been significantly streamlined. This change makes instructions clearer and shorter, saving processing power when the agent is used.

Contributor emails are now correctly linked to their project accounts. This ensures that contributions to specific parts of the project, like the Ramp Router and Nebius Token Factory, are accurately attributed to the right individuals.

We've added support for Ramp Router, a new service that intelligently directs your requests to different AI models. This allows for more flexible and cost-effective use of AI by automatically choosing the best model and providing backups.

Hermes now supports the Nebius Token Factory as a way to use various open-source AI models. This integration allows you to easily access models like Qwen, DeepSeek, and Llama through a single provider.

The project now supports a new way to use Tencent's AI models for generating responses, making it easier to access their latest large-context models. This update also sets a more capable version of Tencent's AI as the default.

The instructions for using the Hermes Agent's skills have been clarified and made more accurate. This removes repetitive text and ensures that guidance about the "hermes-agent" skill only appears when it's actually available.

Tasks can now be organized into nested subtasks, making it easier to manage complex projects and break down work. This change allows the agent to better plan and execute multi-step tasks by visually representing their hierarchy.

When a process that restores a skill from a backup fails, the system now ensures that neither the skill nor its backup copies are accidentally deleted. This prevents data loss and makes it easier to recover from failed restorations.

The default personality for new Hermes Agent users has been updated to significantly reduce overly long responses by prioritizing conciseness and clearly stating what not to do, like adding filler or repeating the request. This change directly addresses common user complaints about verbosity and aims to make replies more efficient and relevant to the original question or task.

Hermes Agent can now update even if some Windows services are temporarily paused, preventing update failures caused by unrelated background processes. This change ensures that updates only fail if a critical service for the agent itself is in a problematic state.

The system prompt now includes two lines to accurately track conversations that last multiple days. The first line shows the original start date of the conversation, and the second line shows the current date, helping the agent understand the passage of time.

The `/busy` command is now available on platforms like Telegram and Discord, not just the command line. This allows users on these platforms to better manage their agent's status and queue.

The `/btw` command now understands the current conversation, allowing it to answer your quick questions without interruption. The `/background` command has been retired, and `/bg` is now the preferred way to run background tasks.

The project's code formatting has been automatically improved, making it cleaner and easier to read. This update ensures consistency across the codebase without requiring manual intervention.

When using Hermes with Slack, link previews will now appear for streamed replies, even if the link is added after the initial message. This change ensures you get the full context from Slack messages as intended.

The `/btw` feature now uses the same caching mechanism as the background review, meaning it can access the full conversation history for answers. This provides more accurate responses by avoiding truncated context and leverages existing efficient caching to keep costs down.

We fixed an issue where disabling reasoning would incorrectly send a setting specific to Ollama to other AI services, causing requests to fail. Now, this setting is only sent when communicating with Ollama, ensuring broader compatibility.

If the Hermes Agent fails to compress a long conversation for a coding assistant, it will now pause before trying again to avoid getting stuck in a loop. This prevents long waits and repeated errors during coding sessions.

Hermes Agent can now correctly use newer models recommended by Nous Portal, even if they haven't been officially added to its list yet. This means users will have access to more up-to-date AI models without interruption.