Changelog

This update fixes an issue where the Hermes Agent would incorrectly display your home directory instead of your project's working directory in its system prompt. This now ensures the prompt accurately reflects your current project context, making it easier to manage your agent's environment.

The agent now correctly formats messages when communicating with NVIDIA's NIM service, preventing errors caused by extra information in tool results. This ensures smoother operation when using NVIDIA models.

When updating Hermes on macOS, the agent will now properly restart if it was previously removed from running services. This ensures the agent is always running after an update, preventing potential downtime.

The floating hearts that appear when you send certain reactions can now be turned off. This gives you more control over the visual effects in the desktop app.

Context menus in modal dialogs will now correctly appear within the dialog window. This prevents issues where actions like pasting would happen outside the intended area, ensuring a smoother user experience.

The agent's display now correctly shows the agent's thinking process without becoming unreadable when you switch away from it. Additionally, if the agent has trouble seeing other application windows, it will now provide a clear reason why it failed, making troubleshooting easier.

The project's code formatting is now automatically corrected for better consistency. This update ensures that code adheres to standard styles without manual intervention, making it easier for everyone to read and contribute.

Security checks are now more robust, preventing accidental exposure of sensitive files. The system will also warn you if these checks are skipped for certain installation locations, and documentation has been updated to reflect these changes and provide recovery steps for locked systems.

Codex requests now connect faster by trying both IPv6 and IPv4 addresses at the same time, picking whichever responds first. This prevents the agent from getting stuck on networks where IPv6 is advertised but doesn't actually work.

Long Signal messages will now be sent in multiple parts, ensuring they are not cut off. This fix preserves message formatting like bold or italic text across these parts.

On Windows, the right-click menu's "Paste" option will now work correctly even if the system thinks the clipboard is empty. This means you can paste text more reliably without the option being unexpectedly grayed out.

The user interface zoom level will now remain consistent when navigating between different views within the desktop application. Previously, switching views could reset the zoom to 100%, requiring manual adjustment.

The way Hermes replays missed events after a disconnection has been fixed, ensuring all events are now delivered. This prevents data loss and ensures you always see the most up-to-date information.

Third-party sandboxes can now connect to Hermes as terminal backends without modifying the core project. This makes it easier for external services to integrate and function as a terminal environment.

Browser snapshots now store full content on disk instead of using costly LLM summaries, saving money and making them more reliable. The outdated "web extract" LLM option has been removed from settings.

The system prompt now correctly reflects only the tools your agent can actually use, preventing confusion. Additionally, the `hermes-agent` tool is now always installed, ensuring the agent can always help manage itself.

Tests that check how groups restart are now more reliable, preventing them from failing unexpectedly in our automated checks. This ensures we can trust our testing process and catch real issues more effectively.

The project's automated code formatting tool can now fix many code style issues automatically, making the codebase cleaner and more consistent with less manual effort. This helps ensure the code is easier to read and maintain for everyone working on it.

The system now correctly handles cases where a model attempts to use a screenshot by rewriting its file path in a way that the system doesn't recognize. This ensures that the intended screenshots are properly attached and delivered to you.

The agent now correctly hides instructions for memory features that aren't actually available, preventing confusion. This makes the agent more reliable by ensuring it only talks about the tools it can use.

Hermes Agent now prevents malicious symlinks from redirecting sensitive browser snapshot data to unintended locations. This change ensures that snapshot files are written securely, protecting your data from potential corruption or theft.

Tests on desktop are now more reliable and won't fail randomly due to an unrelated mock issue. This prevents unrelated code changes from causing CI failures.

You can now adjust how much text the browser tool keeps when taking snapshots of web pages, making it easier to capture either more detail for long pages or less for quicker summaries. This helps tailor the tool's memory usage and the amount of captured information to your specific needs.

Telegram messages will now retry sending for up to 15 seconds if the connection briefly drops, preventing them from getting stuck for hours until the system restarts. This ensures more reliable message delivery during minor network interruptions.

Teams connect calls will no longer unexpectedly crash if the Microsoft Teams SDK isn't fully installed. This makes the Hermes Agent more stable when interacting with the Teams platform, even in cases where only parts of the SDK are present.

The system now shows why it switched to a different model when the primary one failed, and it remembers the full path taken through fallback models. This helps users understand issues with model providers and ensures they know when the original model is back in use.

Attachments created in Docker containers for specific sessions will now be delivered correctly. Previously, these attachments would silently disappear because the system was looking in the wrong place for them.

Skills that are being reviewed can now be successfully updated. Previously, changes to skills would be rejected even after they were viewed, preventing any updates from being applied.

Hermes can now connect to 44 additional services like Canva and Dropbox with a single command. These connections have been streamlined to remove unnecessary features, making them safer and more efficient for you to use.

Hermes now uses a more specific way to refer to user data in OpenViking, preventing errors when saving or loading memories. This change ensures compatibility with future OpenViking updates and works across all user permission levels.

The background review process now correctly understands when it needs to check a skill before making changes. This prevents unnecessary errors and wasted attempts when editing or updating existing skills.

The Hermes Agent can now correctly send screenshots when they are generated by background tasks or scheduled jobs, and it also handles corrupted or incomplete data from tools more reliably to prevent errors. This means you'll get your attachments more often and the agent will be more stable when processing tool information.

The curator can now properly save changes by ensuring it checks for existing content before attempting to write. This fix prevents previously hidden errors where the curator appeared to do nothing when it was actually unable to save its work.

The documentation for the Gemini agent has been updated to use the latest Gemini models, making it more current and aligned with recent advancements. This change ensures users have access to the most up-to-date and capable Gemini models when interacting with the agent.

Persistent Docker containers now correctly stay shared per profile, ensuring one container for your entire profile's usage. This fixes an issue where containers were splitting per chat, which could lead to problems like missing media attachments.

The "screenshot my screen" command now correctly captures everything visible on your display, not just the desktop wallpaper. This ensures you get a true representation of your screen when taking a full screenshot.

The Hermes Agent now provides clearer machine-readable outcomes for approval steps, making it easier to understand why an approval failed. Additionally, interactive password prompts for sudo commands no longer count against task deadlines, improving flexibility and preventing unexpected timeouts.

Teams meeting transcripts can now be successfully downloaded by correctly requesting the VTT format, which is what the Microsoft Graph API expects. Previously, the agent incorrectly asked for JSON, causing downloads to fail.

Expired promotional tools for FLUX 3 video generation have been removed. Paying subscribers can still access FLUX 3 video through the video generation feature.

Repeated web searches and extractions within a 20-minute window will now only count as a single vendor request. This helps reduce costs by avoiding redundant scraping and searching.

Trusted profiles can now share a single persistent Docker container by setting a common key, which helps save resources. If no key is set, each profile will continue to use its own separate container.

We've added a way for memory providers to ensure important conversation history is saved permanently before it's shortened. This prevents accidental loss of data, making our conversation memory more reliable.

The Hermes TUI no longer unnecessarily reinstalls its dependencies every time it launches. This makes the TUI faster to start up by correctly identifying when dependencies are already installed.

The Sessions list in the desktop app will now correctly show your recent conversations even after switching between different profiles or remote connections. This ensures you always see the right chats, regardless of which backend you're currently using.

The system prompt has been cleaned up by removing unnecessary subscription information. This makes the agent more efficient by reducing the amount of text it processes on every request.

Windows users will now get faster automatic updates because Hermes won't pause for long waits caused by the cua-driver installer. Instead, it will ask you to manually upgrade the driver when needed.

The Atlassian integration is fixed and now works correctly, and support for Grafana Cloud has been added to the system. This means you can now use Atlassian tools through Hermes Agent again and also connect to Grafana Cloud.

The Cronjobs section of the desktop app now correctly displays options for every bot instead of showing an "unavailable" message. This allows users to create scheduled tasks for any bot.

The Cronjobs (Routines) pane will now correctly display your scheduled tasks even if the bot list hasn't loaded yet. This change ensures that the pane updates itself automatically once the bot list is available, so you don't miss any important information.

Hermes Desktop will no longer ask for your macOS keychain password every time you open the app. You can now choose to encrypt your saved gateway secrets using the OS keychain if you want to.

The Routines pane no longer crashes when it receives an incomplete error message from the bot. Instead, it will now display a helpful error message, allowing you to see what went wrong.

Hermes now remembers and suppresses repeated cron job failures that have the same error. You can also manually acknowledge specific error types to stop them from repeatedly alerting you.

The New Cronjob dialog now correctly displays the names of bots when scheduling tasks, instead of showing "[object Object]". This makes it easier to select the right bot for your automated tasks.

Tests in our desktop environment now pass reliably because we've fixed how timers are managed when the edit composer is closed. This prevents lingering tasks from causing problems in later tests.

The tool search feature now correctly handles multiple tool requests simultaneously, preventing errors and ensuring all available tools are found. This makes the agent more reliable and efficient when looking for and using tools.

You can now right-click a tab in the in-app browser to open that webpage in your computer's default browser. This makes it easier to access content from the app in your preferred external browser.

Hermes Agent updates on Windows will now work even if the gateway is running. This is because the update process can now correctly identify and ignore the gateway, so you no longer need to manually stop it before updating.

The documentation now highlights Browser Use Cloud as the primary cloud browser option. This change clarifies its capabilities and reflects how Hermes Agent prioritizes it when automatically selecting a cloud provider.

macOS users no longer need to manually re-grant permissions after updates. A new command automatically sets up a secure identity so that critical permissions like Full Disk Access stick permanently. This simplifies the process and ensures ongoing functionality without user intervention.

The in-app browser can now be opened in its own separate window, making it easier to multitask. When you close this window or choose to pop it back in, it will return to the page you were viewing.

The project's code formatting is now automatically corrected when necessary. This ensures a consistent and clean codebase without requiring manual effort.

The tool search is now smarter, allowing you to find more tools with fewer searches and understanding variations in your wording. It also retrieves tool information more reliably, even if some requested tools are not found.