A running log of what's shipping in Hermes Agent — merged pull requests from NousResearch/hermes-agent, browsable one day at a time with plain-English AI summaries of what changed. Checked every 30 minutes.
Last checked 30 Aug 2026, 02:20 UTC
Hermes now better handles conversations where a stored summary is combined with a new live request. This means retrying or undoing actions will correctly target only your latest message, preserving the conversation history.
The conversation display now stays steady while you read it, preventing unexpected jumps as responses finish. This makes it easier to follow along with the conversation without the text moving around.
Custom integrations can now choose whether to use caching for prompts, making them more flexible. This fixes a previous limitation where custom setups couldn't enable caching without complex workarounds.
The Files panel now correctly displays files from the currently active connection, preventing it from showing outdated or incorrect information from other connections. This ensures a more reliable and accurate file browsing experience when you have multiple connections set up.
Jobs that should only run once will now clearly tell you if you try to run them again instead of silently disappearing. You can now explicitly re-activate these jobs if needed.
We fixed an issue where certain chat types were not respecting their configured access rules, causing messages to be incorrectly blocked. This also ensures that a broad access setting in one profile doesn't accidentally grant access in others.
The terminal approval now correctly distinguishes between commands that look dangerous and just mentioning dangerous commands in text. This means Hermes Agent won't accidentally block you from running commands when you're only discussing them, while still protecting against actual destructive actions.
The Hermes configuration update process now correctly reports only the settings it actually saves. Previously, it falsely claimed to have added a setting that was never written to disk, leading to confusion.
The Hermes command-line tool now starts faster by correctly recognizing the `--reasoning` flag, preventing it from unnecessarily slowing down by searching for plugins on every use. This change means you'll see results from Hermes more quickly.
Device pairing now works correctly when Hermes is set up in certain ways. This fixes an issue where pairing codes were being saved in the wrong location, preventing them from being found when needed for approval.
Hermes Agent now correctly identifies the full context window size for local AI models, ensuring they are used to their maximum potential. This fix prevents the agent from mistakenly using the output limit as the context window, improving performance for large-context models.
The batch runner now correctly remembers and skips samples that were previously discarded because they were missing reasoning. This prevents them from being re-processed and incurring unnecessary costs when you resume a batch run.
The Hermes Agent now correctly reports failures when it can't verify that all running gateways have been updated. This prevents the update process from appearing successful when there's a problem communicating with gateways.
On Windows, Hermes will now correctly return control of your command prompt after an update is finished. This prevents the update process from freezing your console.
The Hermes Agent can now reliably send messages on Windows systems and in environments where the Hermes executable isn't automatically found. This fixes issues that previously caused errors when dealing with specific Windows file paths or when running Hermes in restricted environments, ensuring smoother operation across different setups.
When your chat session automatically restarts in the dashboard, you will now be taken directly to the latest messages, making it easier to pick up where you left off. This change improves the user experience by ensuring the most relevant content is visible immediately upon session resumption.
Tests for the Skills view are now more reliable on slower systems, preventing them from failing unexpectedly. This makes the development process smoother by reducing test flakiness.
The Hermes desktop app will now correctly retry connecting to the remote backend if it gets stuck, instead of freezing and requiring a restart. This ensures a more stable and reliable connection experience.
Branching a session now correctly opens it in the main part of your workspace, so you can immediately see and interact with it. This fixes an issue where branched sessions were only appearing in a hidden sidebar, making them seem lost.
Group chats on desktop will no longer crash after a cloud connection is removed. This fix ensures that group chat members are properly handled when a connection is deleted, making the app more stable.
The workspace pane in the desktop app will now remain functional even if some parts of the application are missing due to an incomplete installation. This prevents crashes and allows you to still see your code differences.
This update prevents certain recurring background tasks from running indefinitely, which could previously cause performance issues. Now, these tasks will automatically stop after a set number of repetitions, ensuring smoother operation.
Hermes Agent now logs a warning when a scheduled task is skipped because it was recently re-armed. This change improves visibility by informing operators when a task is intentionally paused and not run.
This update prevents incorrect API keys from accidentally hiding valid ones, ensuring you can connect to services reliably. Now, if an API key is entered incorrectly, a warning will appear and a correct key will be used instead.
The pricing for very long Gemini prompts has been corrected to accurately reflect Google's tiered rates, ensuring users are billed appropriately for extensive usage. This fix matters because it prevents undercharging on the very sessions where prompt length and cost are most significant.
The bot relay now keeps a single connection open while active instead of creating and closing new ones every few seconds, significantly reducing log messages. This change improves efficiency and prevents unnecessary network activity.
Bots in group chats on the desktop app will now respond correctly. This fixes an issue where the bot's connection was being unexpectedly closed, causing it to go silent.
The unread count on the desktop application's title bar no longer incorrectly includes background automated tasks, so it will now accurately reflect only your unread conversations. This makes it easier to see when you have new messages to respond to.
When you open multiple files, they now appear as tabs within a single preview window instead of creating separate side-by-side panes. This makes it easier to manage and view all your open files without cluttering your screen.
Hermes now remembers the exact time events occurred, making its recollections more accurate and useful. This improvement ensures that when you ask Hermes to recall something, it knows precisely when it happened, not just that it happened.
Keyboard navigation now clearly shows which element is selected, making it easier for users who navigate with their keyboard to use the app. This improves accessibility by restoring visible focus indicators that were previously hidden.
The desktop app will now automatically reconnect if the remote gateway restarts, preventing it from freezing and requiring a manual force quit. This makes updates to the remote gateway a smoother experience for users.
The update process now correctly handles users' preferred browsers, preventing unexpected browser windows from opening and ensuring no leftover windows remain after an update. This makes the update experience smoother and avoids potential privacy issues with certain browsers.
Scheduled tasks for desktop profiles that were not actively being used will now run as expected. This change ensures that background jobs continue to execute even when a specific profile is "sleeping."
The way Hermes checks if data has been adopted by another backend has been improved to catch more potential issues. This makes sure your data is less likely to be lost if multiple systems try to access it simultaneously.
The agent now correctly avoids making its installation directory and its subfolders too restrictive. This prevents the agent from locking itself out and failing to run after updates.
The Hermes setup command now gracefully handles cases where it can't properly interact with your terminal, preventing crashes and allowing you to continue using it. This makes the setup process more reliable, especially in unusual terminal environments.
Cron jobs that are not set up to do anything will no longer cause problems or run endlessly. The tool that manages these jobs will also no longer accidentally clear a job's name if you try to update it with a blank name.
The project now correctly credits a new contributor, @cycorld, for their work. This is important to ensure all contributions are properly recognized before a related code change is merged.
The gateway now prevents accidental restarts of other Hermes profiles, which could previously happen if an old configuration file incorrectly pointed to a running gateway. This change ensures that the `--replace` command only affects the intended gateway, preventing unintended disruptions.
Keyboard input is now handled more reliably in the setup menus, so special keys like Escape and Ctrl+C work as expected. You can also use the left arrow key to go back through your previous choices, making the setup process smoother.
This change prevents the agent from freezing for extended periods when processing large responses from the Codex API. By sending certain data directly, it bypasses a slow internal check within the OpenAI software library.
The `/retry` and `/undo` commands now correctly resend your last message after the conversation context has been shortened. This ensures that the agent uses the most recent instruction you provided, rather than an older one.
The update now ensures that the Windows update screen opens in your preferred web browser, not just Edge. It also prevents a temporary browser profile from being left behind, cleaning up your system after each update.
Hermes Agent now automatically reconnects to its server if the connection unexpectedly drops, such as when your computer sleeps. This ensures that the desktop and TUI interfaces remain usable and don't get stuck after network interruptions.
A problem where the assistant would sometimes speak replies twice has been fixed. A new test has been added to ensure this fix stays in place, preventing the issue from reappearing in the future.
The agent now correctly warns when you schedule a job but the necessary gateway isn't running to execute it. This prevents jobs from appearing scheduled when they are actually unable to run, providing clearer feedback.
The Telegram integration has been updated to use a more robust and shared deadline system. This change improves how tasks with deadlines are managed, ensuring that cancelled tasks are properly cleaned up and preventing potential issues like resource leaks.
Tool call timeouts in the MCP component are now handled more consistently. This change ensures that tool calls are managed by a single, unified timeout setting, making it easier to configure and preventing unexpected timeouts.
New chat sessions started from the Home screen will now correctly remain separate from any previous project you were working on. This prevents your new chats from accidentally showing information or being linked to an old project's files.
The Hermes Agent now provides clearer messages and a specific exit code when the port it needs is already in use. This change helps users more easily identify and resolve port conflicts, preventing confusion with other errors.
The desktop application will no longer restart a healthy backend when a probe fails, preventing unnecessary interruptions. Additionally, any errors encountered when starting the backend will now include detailed output, making it easier to diagnose and fix problems.
The HUD window now correctly floats and stays on top of other windows when using Hyprland, a specific window manager. This ensures the HUD remains accessible and draggable, improving your user experience.
The project's code formatting is now automatically corrected for consistency, making it easier to read and maintain. This update streamlines the development process by handling routine style fixes automatically.
The way the desktop HUD (Heads-Up Display) interacts with your operating system, like how it moves, ignores mouse clicks, or stays on top, has been simplified. This change makes the HUD more reliable and consistent across different desktop environments without altering its behavior for most users.
Hermes now correctly identifies itself when communicating with OpenAI's Codex service, preventing impersonation and improving compatibility. This change ensures Hermes is recognized properly by OpenAI's systems.
The system now correctly recognizes special suffixes like ":nitro" and ":floor" when requesting models from OpenRouter. This means you can use these options to fine-tune your requests and ensure they are processed accurately, rather than being rejected or silently changed.
You can now press Cmd+L (or Ctrl+L on Windows/Linux) from anywhere in the app to quickly jump to the message composer. This makes it faster to start typing a reply or new message without needing to navigate around first.
The code formatting and style rules in the JavaScript parts of Hermes Agent have been automatically corrected. This helps keep the project's codebase clean and consistent.
You can now open multiple tabs in the in-app browser, similar to a real web browser, allowing you to view different pages simultaneously. The address bar also shows improved loading progress and holds your entered address until the page loads.
When you switch between conversations in Hermes Desktop, questions you were asked might become hidden and unanswerable. This update ensures those questions reappear correctly so you can continue your conversation without interruption.
The project's automated tools now automatically fix code formatting and style issues. This helps keep the code clean and consistent without requiring manual effort.
Hermes can now automatically recover from certain network issues and disconnects more quickly when a connected service becomes unresponsive. This makes Hermes more reliable and less likely to get stuck.
Hermes Agent will now properly detect and replace unresponsive Language Server Protocol clients, preventing the program from freezing and providing faster, clearer error messages when something goes wrong. This makes the agent more robust and easier to debug when language server features are not working as expected.
Hermes Agent now automatically restarts its computer control session if it times out during an action. This prevents the agent from getting stuck and clearly indicates when an action's outcome is uncertain due to the timeout.
Hermes Agent now prevents processes started in the terminal from being left behind when a local command times out. This ensures that all related processes are properly cleaned up, preventing resource leaks.
The Hermes Agent now uses a single, consistent way to stop process trees, making it more reliable and easier to manage when programs need to be terminated. This change helps ensure that all related processes are cleaned up properly.
The Hermes desktop app can now recover lost messages during network interruptions. This prevents data loss for things like truncated tool output or incomplete assistant replies, so you won't have to reload the app as often.
Automatically fixed code formatting issues to improve code quality and consistency. This ensures the codebase remains tidy and easier to maintain.
The titles and headings on important documentation pages have been updated to better match what people search for, making them easier to find. Additionally, the desktop documentation now links directly to the product page for easier access.