Changelog

The Hermes Agent now supports a newer version of NeMo Relay, allowing for more flexible and dynamic activation of plugins. This change enhances observability by enabling Hermes to manage LLM and tool middleware more effectively.

Commands in the terminal will now run in the correct directory, even if a different session previously changed directories. This prevents commands from accidentally running in the wrong project folder.

The Hermes Agent now correctly handles interactions within Slack, ensuring that actions taken in one workspace don't interfere with another. This makes the Slack integration more reliable and prevents confusion when using multiple Slack workspaces.

The user interface for displaying assistant messages has been updated to use newer built-in features for smoother and more efficient text streaming. This upgrade also fixes a potential issue where the app might briefly go blank during session changes or reconnections.

This update fixes an issue where the "Summarizing thread" message would incorrectly persist after a session resumed activity. It also ensures that timers for multiple active sessions now correctly track elapsed time independently when switching between them.

Empty user messages are now removed before being sent to AI providers. This prevents errors and ensures consistent responses from different AI models.

Messages in Telegram will now only use a special table format when you explicitly enable it. This prevents tables from appearing too large or oddly on mobile devices by default.

The dashboard now remembers which service issued your login. This prevents you from being unexpectedly logged out when using multiple authentication methods.

A test that was sometimes unreliable has been removed to improve the stability of our testing process. This change ensures that our automated checks are more trustworthy.

You can now set specific reasoning effort levels for different AI models directly in the configuration file. This makes it easier to fine-tune how each model thinks without manually changing global settings.

Hermes can now run safe tool calls at the same time as other tool calls that need to run one after another, making the agent faster. This prevents a single slow or incompatible tool from delaying all other independent operations.

Blender can now be easily installed and used with Hermes, providing a curated set of essential tools for 3D scene manipulation. This update also ensures that features like telemetry are disabled by default for privacy, requiring explicit user opt-in.

You can now use the Upstage Solar AI model with Hermes. This allows you to leverage a new powerful AI option for your tasks and benefit from its unique reasoning capabilities.

The Hermes Agent now retries responses that send no data, similar to how it handles connection errors. This change also ensures that issues with specific AI providers like Anthropic are handled more consistently and with clearer error messages, preventing users from mistakenly troubleshooting network problems.

Previously, certain types of content from MCP tools were hidden. Now, these resources are properly displayed, making tool outputs more complete. This also ensures error messages are more informative when issues occur.

The "Create Backup" button in the dashboard now functions correctly by properly passing the backup file location to the command. This fixes an issue where backups were failing due to an incorrect command format.

The Telegram integration now correctly connects on Python 3.13. This resolves an issue where a recent change in Python made it impossible for Telegram to poll for new messages.

The Nous AI model provider has been updated to use the correct web address, fixing an issue where models would time out because the old address was no longer working. This ensures you can reliably connect to and use Nous AI models.

Tools provided by the Hermes tools server now correctly describe what information they expect, allowing them to properly receive and use arguments. This fix makes the entire set of tools usable from other programs.

The Hermes Agent now prevents excessively large terminal output from crashing the system. This change ensures that even very chatty programs will not overwhelm the agent with too much data.

Hermes CLI sessions will now correctly identify and process their own asynchronous tasks, preventing them from accidentally taking over tasks belonging to other, older sessions. This ensures that tasks are handled by the intended session, improving reliability.

The system will now more reliably detect and recover from hangs when processing large requests. Previously, a watchdog meant to catch these hangs was turned off for large requests, causing them to fail much later or incorrectly.

You can now control how much "thinking" each side task, like image description or compression, performs independently. This allows you to make these tasks faster and cheaper without affecting the main chat's intelligence.

You can now easily turn the subtle statue background in your chat on or off directly from the Appearance settings. This gives you more control over the look of your chat window.

The way Hermes Agent saves its configuration has been improved to prevent accidental deletion of important settings like platform credentials. This update ensures that when parts of the configuration are saved, unrelated sections are no longer lost.

Database checkpoints are now less disruptive, preventing potential data corruption on large databases. This change ensures that the system can safely write data without causing issues with message indexes.

We've fixed an issue where the Hermes Agent would continue thinking even when told not to, specifically when using Ollama models. This change ensures that when you set the agent to "reasoning_effort=none," it correctly stops its thinking process, saving time and resources.

Messages in the desktop app will no longer contain unwanted text fragments caused by copying and pasting. This ensures that your conversation history is clean and accurate.

Desktop cron jobs will now start successfully even if the software was updated while they were running. This prevents cron jobs from failing due to version mismatches and ensures they can use a fallback client if the preferred one isn't available after an update.

The assistant's thinking status in Slack will now correctly clear when you stop an agent, and the typing indicator will resume immediately after you respond to a clarification prompt, preventing it from getting stuck. This ensures a smoother and more responsive interaction experience within Slack.

You can now control how deeply each part of the MoA (Mixture of Advisors) system thinks. This allows for more precise control over reasoning effort for each advisor and the aggregator individually.

The system can now handle cases where tools don't provide arguments without crashing the entire conversation. This ensures conversations continue smoothly even if a tool's input is unexpectedly missing.

The default model used when none is selected has been changed to a less expensive option. This prevents users from accidentally being charged for premium models they didn't choose.

The system now reuses a key checker for verification requests, preventing redundant fetches and avoiding issues where requests were being rejected or timing out. This makes the verification process more efficient and reliable, especially when many requests happen at once.

The agent can now understand information about the surrounding conversation context when it needs to act on messages. This allows bots to respond more intelligently to requests that reference earlier parts of a chat.

The way the assistant displays prompts has been improved to prevent important clarification questions from being hidden. Now, these prompts will always be fully visible, even when there are many tools being used.

The Hermes Agent can now better identify and remove common writing flaws by adding new detection patterns for things like forced metaphors and dramatic sentence endings. Additionally, the agent's own explanations have been improved to avoid the very writing issues it's designed to find, making it a clearer teacher.

The MoA aggregator can now correctly use global reasoning settings when a specific model isn't set for it. This ensures that the aggregator uses appropriate reasoning levels, improving its performance and reliability.

Hermes Agent now prevents OpenAI Codex requests from running indefinitely by setting a maximum time limit for each request. This ensures that requests that get stuck will eventually time out, preventing sessions from hanging.

The Hermes Agent now correctly recognizes when a reasoning-only response from certain providers is complete, preventing unnecessarily long waits. This update ensures that the agent doesn't get stuck waiting for tasks that have already finished, improving its responsiveness.

The test for handling interruptions in asynchronous operations was made more reliable by increasing internal timeouts, preventing it from failing on busy systems. This ensures the test accurately reflects how Hermes Agent handles interruptions, even under load.

Hermes can now recover its answers when the AI model unexpectedly embeds the final response within its internal thinking process. This fix also prevents the agent from getting stuck in a loop when it can't immediately retry a failed response.

The system now has a clearly designated "default" model that can be easily changed without requiring a new software release. This ensures that if no specific model is requested, a safe and cost-effective option is always used, and simplifies switching the default model in the future.

The display of tool calls in Hermes Agent has been improved to prevent crashes when tools provide unexpected data types. This ensures a smoother and more reliable user experience when interacting with tools.

The Hermes Agent now tells you why it's taking a while to get a response. Instead of just showing a spinner, it will explain if a provider is slow or overloaded, making it clearer when there are delays. This helps you understand what's happening when the agent is working.

Tests for the Assistant UI have been updated to correctly find the shared tap library, even when it's installed in a different location than before. This ensures that the tests accurately reflect how the project is built and that the correct version of the tap library is being used across all Assistant UI components.

A test for how Hermes handles responses from its Codex backend was failing because recent code changes altered the expected behavior. This update adjusts the test to correctly account for how the backend's identity should be recognized, ensuring the test accurately reflects the system's intended operation without affecting production code.

Web search results in tool summaries are now displayed as direct URLs instead of generic placeholders. This change makes it easier to see where information came from without needing to click on a link.

Now, custom configuration details you add for auxiliary calls made through Anthropic will be correctly sent to the Anthropic API. This ensures that specific instructions and metadata you provide are respected by the system, preventing them from being lost.

We've updated how some parts of Hermes Agent communicate with AI models. This change ensures that all requests correctly use configuration settings, leading to more consistent and reliable AI-powered features.