NousResearch/hermes-agent
46.5
Weak · 25 October 2026
1457.8k
lines of production code
Python
with TypeScript
3
measurements over time
What this system is
Hermes is a cross-platform AI agent platform that provides a Python backend gateway alongside native desktop, CLI, and TUI clients. It enables autonomous coding and web automation through a pluggable architecture that supports diverse model providers, secret managers, and browser backends. The system includes a comprehensive evaluation harness for benchmarking agent performance and a rich desktop interface with features like floating pets, Kanban boards, and session management.
Features
Add TUI benchmarking scripts and build tooling
The ui-tui area now includes a suite of benchmarking scripts to measure performance characteristics of the terminal interface. These include \bench-history-scroll.tsx\ and \bench-native-history.tsx\ for comparing virtualized versus native transcript scrolling, and \bench-streaming-md.tsx\ for evaluating streamed markdown rendering strategies. Additionally, build and utility tooling has been added, such as \i18n-export-en.ts\ and \i18n-keys.ts\ for managing translation catalogs, and \clipboard\_image.py\ for handling image pasting dependencies.
ui-tui · high confidence
Add opt-in F12 DevTools shortcut toggle
The desktop app now includes a configurable setting to disable the F12 keyboard shortcut that opens the developer tools. This feature is opt-in, allowing users to prevent accidental activation of the DevTools window.
apps/desktop/electron (part 3 of 6) · high confidence
Add opt-in Radio plugin with station playback and artist profile links
A new Radio plugin is available in Hermes Desktop, disabled by default and enabled via Capabilities → Plugins. It adds a status-bar widget with a braille waveform, play/pause, and next-station controls, plus a station picker that includes pinned presets (Nightride FM, Radio Paradise, EVE Radio), search via Radio Browser, and volume control. Playback starts only when you press Play; pausing releases the stream, and resuming returns to live. Artist names in the now-playing metadata link directly to artist profiles (Apple Music for unambiguous names, MusicBrainz as a fallback), with results cached for a day and requests paced across windows. The plugin uses plugin-scoped local storage for pins, station, and volume, respects reduced-motion preferences, and performs no work when disabled.
apps/desktop/src/plugins/radio · high confidence
Add session import view for foreign coding-agent transcripts
Users can now import sessions from external coding agents (Claude Code and Codex) directly within the desktop app. This change introduces a new sidebar view that lists, previews, and imports these foreign sessions, handling the API communication and UI state for the import process.
apps/desktop/src/app/session-import · high confidence
Added archive confirmation dialog for skills with profile scoping
A new confirmation dialog allows users to archive learned skills, featuring optimistic UI updates that roll back on failure. The dialog supports optional profile scoping, enabling users to archive skills against a specific profile's backend rather than just the app-wide active profile, and provides localized success and error notifications.
apps/desktop/src/app/learning · high confidence
Added egress proxy integration module
Added the \agent.proxy\_sources\ package to provide a unified location for egress proxy integrations, specifically serving as the entry point for the \iron\_proxy\ module which wraps the ironsh/iron-proxy tool.
_agent/proxy\sources · high confidence
Added plugin-host readiness audit tool
A new CLI tool (\evals/plugin\_isolation/audit\_catalog.py\) has been added to evaluate the compatibility of third-party plugins with the per-profile plugin host. This tool fetches plugins from the catalog at their pinned commits and runs an isolation audit to determine which plugins can run in the isolated host environment versus those that require in-process loading, providing a summary report of readiness and specific blockers.
_evals/plugin\isolation · high confidence
Agent preview now uses durable element handles and delta updates
The in-app browser preview now provides a more efficient and stable experience for the automation agent. Instead of sending the entire page inventory on every turn, the agent now receives durable, human-readable handles (like \btn-sign-in\) that persist across page changes, and subsequent updates are sent as lightweight deltas showing only what changed. This significantly reduces context usage and prevents errors caused by elements shifting during re-renders. Additionally, the preview now includes a visual overlay that highlights the elements the agent is interacting with, allowing you to see exactly what the agent is doing in real-time.
apps/desktop/src/lib/preview-act · high confidence
Arabic locale support added to the Desktop app
The Desktop application now includes a complete Arabic (ar) translation catalog. This adds Arabic language support for the user interface, including the command center, settings, skills and plugins catalog, session management, and onboarding flows.
apps/desktop/src/i18n · high confidence
Billing settings page revamp with in-app plan management and auto-reload controls
The billing settings page has been completely redesigned to support in-app subscription changes, including a new current-plan card that displays tier details and allows users to preview or schedule downgrades (with an undo option). The page now features a dedicated auto-reload row for configuring threshold and reload amounts, a revamped top-up interface with input validation and preset buttons, and a new in-app plans catalog view for browsing and selecting tiers. Error handling has been standardized with a new refusal resolution system that maps API errors to user-friendly messages and actions (such as retrying, stepping up, or redirecting to the portal). The implementation includes a new billing API layer with idempotency support for charges and comprehensive test coverage for the new components and logic.
apps/desktop/src/app/settings/billing · high confidence
Bot Mode plugin bundled as default-on with comprehensive test coverage
The hermes-bots plugin is now bundled as a built-in, default-on feature in the Desktop app. This release introduces a new avatar system with deterministic blob faces, an image generation tab, and device upload, alongside a fix to extend the avatar image generation timeout to 90 seconds. It also corrects bot row click behavior to consistently open the canonical Bot Chat instead of re-opening closed tabs, and ensures that clicking a bot row lands on the correct session even if it was previously compressed. The change includes extensive test coverage for avatar rendering, bot deletion, and session focus identity.
apps/desktop/src/plugins/hermes-bots · high confidence
CJK-optimized search index for session messages
A new native SQLite FTS5 extension (fts5\_cjk) is added to improve search performance for Chinese, Japanese, and Korean text. The extension implements a 'cjk\_unicode61' tokenizer that breaks CJK character runs into overlapping bigrams, preventing short queries from falling back to slow full-table scans. When the extension is installed, the application automatically creates a 'messages\_fts\_cjk' full-text index on the v23 external-content layout, excluding tool rows. Users can disable this feature by setting 'sessions.cjk\_fts: false' in their config file.
native · high confidence
Centralized tool and application discovery across operating systems
The resolver module now provides a unified system for locating installed tools and applications by defining OS-specific directory tables for ecosystems like Homebrew, Rust, Node, and Python, as well as platform-specific locations for applications (e.g., macOS /Applications, Windows Program Files) and Hermes-specific data directories. This change consolidates how the platform probes for binaries and inspects their state, ensuring consistent discovery on Windows, macOS, and Linux without scattering path logic throughout the codebase.
_hermes\platform/resolver · high confidence
Clarify cards now display question/answer pairs and handle undelivered requests
The Clarify tool now renders a settled view that explicitly shows the question and the user's answer (or a 'skipped'/'no answer' status) after submission, replacing the previous collapsed fallback. It also introduces an 'undelivered' state that detects when a clarification request fails to reach the client within a grace period, displaying a warning notice instead of a silent spinner. Additionally, the component now supports a keyboard shortcut (Cmd/Ctrl+Enter) to submit answers and correctly distinguishes between single- and multi-select question shapes in the UI.
apps/desktop/src/components/assistant-ui/clarify · high confidence
Command palette now supports extensible contributions and live highlight previews
The command palette (⌘K) now uses a contribution system that allows other parts of the app to register rows, enabling features like installing VS Code themes from the Marketplace, browsing and adopting pets, and toggling settings directly from the palette. To support this, the palette now correctly tracks the currently highlighted item in uncontrolled mode using a new \HighlightWatcher\ component, ensuring live previews and state updates work reliably.
apps/desktop/src/app/command-palette · high confidence
Cron jobs are now first-class sidebar entities with blueprint support and live sync
Cron jobs are now a primary entity in the desktop sidebar, accessible via the nav rail, and can be created from pre-built Automation Blueprints that seed form fields and adapt delivery targets for the desktop. The interface supports per-job model overrides, multiple delivery targets, and script-only jobs, while providing immediate visual feedback when triggering a run. State is kept in sync via event-driven live broadcasts instead of polling, and the system now correctly gates write access to run transcripts based on scheduler ownership to prevent editing zombie sessions.
apps/desktop/src/app/cron · high confidence
Delegation routing settings: model and profile routes
The Settings page now includes a Delegation Routing section where you can choose a routing mode (Default, Model routes, or Profile routes) and configure the specific routes. In Model routes mode, you can create, edit, and delete named routes that specify a provider, model, reasoning effort, and service tier, with validation to prevent duplicate names and ensure required fields are filled. In Profile routes mode, you can select which profiles subagents may run as, provide descriptions, and set a default profile. The system ensures that a default route or profile is always selected while options exist, and it preserves existing keys (like API keys) during renames. If a plugin-based mode is selected but the backend is unavailable, the UI indicates this and offers to install the Advanced Delegation plugin.
apps/desktop/src/app/settings/delegation · high confidence
Desktop theme system overhaul with per-profile assignments and VS Code marketplace support
The desktop theme engine has been rebuilt to support per-profile appearance assignments, allowing different profiles to maintain independent light/dark modes and skin selections. The system now syncs live skin changes from the Hermes backend, ensuring the desktop UI repaints correctly when a skin is activated via the CLI or agent. Users can now install any VS Code theme from the Marketplace, which are converted into desktop theme families that support both light and dark variants. Accent colors are managed via a new swatch system with a live-preview picker, and custom CSS from user skins is preserved and applied on top of built-in palettes. The Classic Hermes theme is now an explicit, persistent desktop pick rather than being inferred from the stock configuration.
apps/desktop/src/themes · high confidence
Enhanced file and image handling in the desktop chat composer
The desktop chat composer now supports dropping files and folders from the operating system (Finder/Explorer) directly into the chat area, as well as dragging links from browsers. OS file drops are intelligently routed: images are uploaded with local-first previews to prevent UI freezes, while non-image files and folders are attached as inline references (@file, @folder) when the backend can resolve the paths, or uploaded if necessary. The system also distinguishes between native OS drags and internal app drags (like project tree items), ensuring correct handling of line-range references and avoiding duplicate uploads. Additionally, large pastes are converted to text attachments, and image previews are downscaled to maintain performance.
apps/desktop/src/app/chat/hooks · high confidence
Free-tier sign-in flow and notice strip introduced
Hermes Desktop now presents a dedicated sign-in dialog for the free tier, handling the full OAuth device-code flow and displaying a post-sign-in result that details which apps or connectors moved or need reconnection. A new notice strip appears in the composer status bar to offer the free tier to existing users, with logic ensuring only one composer instance displays the notice at a time to avoid duplication.
apps/desktop/src/components/free-tier · high confidence
HUD mode now features a native frost effect, click-through transparency, and improved window management
The desktop HUD has been redesigned to behave like a native Spotlight-style bar. It now supports a native frost (glass) effect behind the band that activates only when the user is typing, avoiding visual artifacts when idle. The window intelligently ignores mouse clicks in empty areas, allowing users to interact with content behind the HUD, while remaining solid over interactive elements like the composer and portaled dialogs. Window movement is handled via a long-press drag on the composer (with immediate Ctrl+drag support on X11) to prevent interference with text selection, and the HUD correctly hands session focus and composer caret back to the main app window upon closing.
apps/desktop/src/app/hud · high confidence
Improved terminal performance and reliability
The integrated terminal now streams agent output live instead of polling, and reuses a single persistent xterm instance across tabs to keep shells alive and avoid layout garbling. Background agent terminals are mirrored as read-only tabs with bounded memory (evicting old processes) and proper re-seeding from snapshots. Clipboard operations (copy/paste) now work reliably via a hidden textarea mirror, and clicking links opens them in the in-app browser. The terminal pane includes a tab rail, respects window visibility to pause rendering, and fixes focus, resize, and tooltip behaviors.
apps/desktop/src/app/right-sidebar/terminal · high confidence
Inline embeds, diagrams, and alerts in assistant messages
The assistant transcript now renders rich inline content directly within messages. Users can view property listings as browsable cards with photo mosaics, see Mermaid diagrams rendered as scalable images with full-screen pan/zoom, and view SVG graphics. GitHub-style alert blocks (Note, Tip, Important, Warning, Caution) are now displayed with distinct icons and colors. Social media posts from X and Instagram appear in isolated, sandboxed iframes, while YouTube, Spotify, and Google Maps embeds load in cross-origin frames. To protect privacy, social and third-party embeds are hidden behind a consent gate by default, requiring the user to explicitly load them.
apps/desktop/src/components/assistant-ui/embeds · high confidence
Installer UI components ported from desktop app
The bootstrap installer now includes standalone UI components (BrandMark, Button, HackeryButton, Loader) ported from the desktop application to ensure visual consistency. The Button component replicates the desktop's shadcn-style styling to match brand defaults without relying on the desktop's theme system. The Loader component implements the 'Fourier Flow' animation curve used in the desktop app, and the BrandMark displays the nous-girl icon in a squircle. These components provide the installer with a unified design language and specific interactive elements like the onboarding 'Begin' button.
apps/bootstrap-installer/src/components · high confidence
Introduce @hermes/ink as a forked, standalone Ink renderer with improved ANSI and mouse handling
The TUI now uses a forked version of the Ink rendering library, distributed as the @hermes/ink package. This fork resolves a critical startup deadlock caused by duplicate Ink instances by deliberately excluding the ink-text-input dependency from the main entry point (consumers must import it via a dedicated subpath). It also corrects ANSI SGR transitions—ensuring bold and dim styles properly reset via SGR 22 to prevent visual drift—and adds a mouse-tracking watchdog to recover from lost DEC modes. The package includes new components like AlternateScreen and RawAnsi, hooks for stdin/stdout, and platform-specific fixes for Windows bidirectional text and legacy Apple Terminal color support.
ui-tui/packages/hermes-ink · high confidence
Introduce ACP adapter for editor integration
Adds the \acp\_adapter\ package, providing the Agent Communication Protocol (ACP) bridge that allows editors (such as Zed) to connect to the Hermes agent. This new component handles authentication detection, session cataloging, slash command dispatch, and content staging, enabling editor clients to interact with the agent via a standardized protocol.
_acp\adapter · high confidence
Introduce Chronos, a NAS-mediated cron provider for scale-to-zero agents
A new cron provider named 'chronos' is now available, allowing hosted agents to delegate job scheduling to the Nous Infrastructure Service (NAS) instead of relying on a local periodic ticker. This enables agents to scale to zero while still firing scheduled jobs via NAS one-shots. The provider includes a dedicated HTTP client for NAS communication, robust handling of identity rejection (falling back to the built-in ticker if NAS credentials are invalid), and secure verification of incoming fire callbacks using cached JWKS clients to prevent rate-limiting issues.
_plugins/cron\providers · high confidence
Introduce Google Meet plugin for transcription, realtime speech, and remote node hosting
The new google\_meet plugin enables the agent to join Google Meet calls, scrape live captions into a transcript, and optionally speak into the meeting in realtime via OpenAI Realtime. It supports two operating modes: local (default) where the Playwright bot runs on the gateway, and remote node, which allows the bot to run on a separate machine (e.g., a user's Mac with a signed-in Chrome profile) while the gateway controls it via WebSocket RPC. Realtime speech requires a virtual audio device (PulseAudio on Linux, BlackHole on macOS) to feed generated audio into the meeting's microphone. The plugin registers five tools (meet\_join, meet\_status, meet\_transcript, meet\_leave, meet\_say) and a CLI (hermes meet) for setup, authentication, and lifecycle management.
_plugins/google\meet · high confidence
Introduce Hermes Desktop native application
Adds the Hermes Desktop app, a native Electron-based chat surface for macOS, Windows, and Linux. This location provides the application's core documentation (AGENTS, DESIGN, BUILDING, README), the shadcn UI component configuration, and the Electron packaging configuration (electron-builder.config.cjs) that defines the installer artifacts, update feeds, and platform-specific build settings.
apps/desktop · high confidence
Introduce Hermes Kanban dashboard plugin and deprecate standalone dispatcher
The Kanban multi-agent collaboration board is now delivered as a dashboard plugin (plugins/kanban/dashboard) that renders the board view, drag-and-drop interactions, and live WebSocket updates directly in the web UI, with styles and logic bundled in dist/index.js and dist/style.css. The plugin SDK integration includes a Checkbox shim for older hosts and a robust i18n fallback so the board remains usable on older dashboards. Additionally, the legacy systemd unit for the kanban dispatcher (hermes-kanban-dispatcher.service) is deprecated in favor of the gateway-embedded dispatcher (controlled by kanban.dispatch\_in\_gateway), with the unit now marked for removal to prevent race conditions.
plugins/kanban · high confidence
Introduce Hermes desktop installer app
Adds a new macOS desktop application named 'Hermes' built with Tauri and Vite. The app serves as a launcher, utilizing a React and Tailwind CSS frontend (configured via vite.config.ts on port 5175) and a Rust backend (indicated by src-tauri in .gitignore). TypeScript compilation targets ES2023, and the project includes standard build artifacts and linting configurations.
apps/bootstrap-installer · high confidence
Introduce Hermes desktop setup installer
A new standalone Tauri-based installer application has been added to bootstrap the Hermes desktop environment. It provides a four-step user flow (Welcome, Progress, Success, Failure) driven by state from the Rust backend, featuring a redesigned UI that inherits the desktop's design system and automatically follows the OS light/dark appearance.
apps/bootstrap-installer/src · high confidence
Introduce Hermes shared metrics observability
Adds a new shared-metrics observability subsystem to the CLI, including a SQLite-backed store for durable aggregation, a consent flow with explicit opt-in/opt-out, and a public-name catalog that collapses user-defined identities to 'custom' to prevent data leakage. This enables bounded telemetry collection for activity, outcomes, and error classes, with daily export capabilities and robust handling of contract drift.
_hermes\cli/observability · high confidence
Introduce LSP integration for post-write semantic diagnostics
Adds a new Language Server Protocol (LSP) layer to the Hermes Agent that runs real language servers (such as pyright, gopls, and TypeScript) as subprocesses to provide semantic diagnostics. The system is gated on git workspace detection to prevent spawning daemons in non-project directories, supports custom server declarations via configuration, and includes a CLI for managing server installation and status. It features a delta-filtering mechanism that compares post-write diagnostics against pre-write baselines to report only new errors, while sanitizing diagnostic fields to prevent prompt injection and ensuring that diagnostic checks do not block file writes.
agent/lsp · high confidence
Introduce Memory Graph visualization with playback, editing, and sharing
The Star Map now features a playable radial timeline that visualizes learned memories and skills as a tilted disk, complete with cinematic node birth animations and a scrubber to replay the graph's growth over time. Users can right-click nodes to edit their content or delete memories and archive skills directly from the map. The visualization supports both dark and light themes with distinct color palettes for memories versus skills, and includes a share-code system that allows users to export their graph as a compact, clipboard-safe string to import into other instances.
apps/desktop/src/app/starmap · high confidence
Introduce Quick Entry global hotkey mini composer
Added a new Quick Entry feature that opens a global hotkey-triggered mini composer window for capturing prompts. This lightweight surface includes an input field, a session-target picker, and logic to handle drafts, submissions, and late relay outcomes, while ensuring theme compatibility and correct shadow rendering on macOS.
apps/desktop/src/app/quick-entry · high confidence
Introduce Tauri-based bootstrap installer with robust update locking and cross-platform script execution
The bootstrap installer has been rewritten in Rust using Tauri to replace the previous Electron-based runner, providing a unified installation and update experience across Windows, macOS, and Linux. This new installer manages the full bootstrap lifecycle—including resolving install scripts from dev checkouts or GitHub, executing them via PowerShell or Bash, and streaming progress events to the frontend—while introducing a robust, cross-process 'update in progress' marker system. This marker system prevents concurrent updates and self-heals from stale locks caused by crashes, ensuring that the desktop application and the installer do not race over the same repository checkout. Additionally, the installer handles platform-specific details such as UTF-8 BOM insertion for PowerShell scripts, ANSI escape stripping for UI logs, and proper process handoff on macOS to avoid duplicate Dock icons.
apps/bootstrap-installer/src-tauri/src · high confidence
Introduce Teams meeting pipeline plugin with operator CLI and Graph-backed processing
A new \teams\_pipeline\ plugin is added, providing an operator-facing CLI (\hermes teams-pipeline\) to inspect, validate, and manage Microsoft Teams meeting processing jobs. The plugin introduces a durable, JSON-backed store for tracking jobs and Graph subscription state, and wires into the Microsoft Graph webhook adapter to automatically schedule pipeline runs upon receiving meeting notifications. It handles transcript and recording retrieval via the Graph API, supports organizer-scoped meeting lookups, and includes outbound delivery to Notion, Linear, and Teams channels.
_plugins/teams\pipeline · high confidence
Introduce composer status stack with coding row, goal indicators, and preview handling
Adds a new status stack component above the chat composer that displays session-scoped status items including a coding row (showing branch, diff counts, and worktree actions), goal indicators, subagent sections, and preview rows. The coding row allows opening the review pane from branch/diff counts and copying the working directory path inline. Preview rows handle remote file artifacts by opening them in the in-app preview pane via the gateway, while local files and remote HTML use the browser bridge. The stack groups items by type (goals, todos, subagents, background processes), respects interface mode tiers (hiding advanced items in Simple mode), and implements polling guards to prevent dead-runtime storms and skip polls for hidden panes.
apps/desktop/src/app/chat/composer/status-stack · high confidence
Introduce disk-cleanup plugin for automatic ephemeral file management
The new disk-cleanup plugin automatically tracks and cleans up ephemeral files (such as test scripts, temporary outputs, and cron logs) created during Hermes Agent sessions. It operates silently via \pre\_tool\_call\, \post\_tool\_call\, and \on\_session\_end\ hooks to identify and manage files without requiring manual agent intervention. Users can manage cleanup via slash commands like \/disk-cleanup status\, \/disk-cleanup dry-run\, and \/disk-cleanup quick\. The plugin enforces strict safety boundaries, ensuring it only operates within \$HERMES\HOME\ and \/tmp/hermes-\\, while protecting critical directories like \logs/\, \memories/\, and user project trees from accidental deletion.
plugins/disk-cleanup · high confidence
Introduce durable cron job scheduling and Kanban work queue
Adds a new cron subsystem for scheduling recurring and one-shot jobs, including a job store, scheduler tick loop, execution ledger, delivery queue, and incident tracking. Introduces a Kanban multi-agent work queue with board persistence, dispatcher, and CLI tools. Includes plugin-contributed automation blueprints, legacy Bot Chat pending record handling, and profile-scoped environment settings for cron.
cron · high confidence
Introduce experimental relay adapter for gateway-to-connector communication
Adds a new \gateway/relay\ package that implements a generic \RelayAdapter\ and \WebSocketRelayTransport\ to bridge the Hermes gateway with an external connector. This feature enables the gateway to front multiple platforms (such as Discord and Slack) over a single WebSocket connection, handling inbound message admission, outbound egress authorization, media upload/download, and interactive prompt approvals. The implementation includes configuration-driven activation via \GATEWAY\_RELAY\_URL\, identity provisioning, and a command manifest that mirrors native Discord slash commands, allowing the gateway to operate as a relayed platform without per-platform gateway code.
gateway/relay · high confidence
Introduce floating pet mascot with roaming, sprite rendering, and hatch visuals
Adds a draggable, always-on-top floating pet mascot to the desktop app. The pet renders pixel-art sprites on a canvas with HiDPI support, scales with Alt+wheel or cursor zoom, and faces inward toward the content. It roams along visible UI surfaces (ledges) with realistic idle behavior (loafing, strolling, hopping) and pauses animations when the window is hidden or inactive. A new egg-hatch animation and star-shower celebration accompany the pet generation flow. The mascot displays a speech bubble showing agent status (working, reviewing, error, waiting) and plugin messages. It persists its position across sessions and refreshes via event-driven live sync with a slow backstop poll for compatibility.
apps/desktop/src/components/pet · high confidence
Introduce floating pet overlay with interactive gestures and click-through
A new pop-out overlay window allows the pet to float freely on the desktop, independent of the main application window. Users can drag the pet to reposition it, use Alt+wheel to zoom it toward the cursor, and interact with it via single-click (opens a mini composer), double-click (toggles the main app window), or shift-click (returns it to the main window). The overlay features click-through behavior, meaning clicks on transparent areas pass through to the underlying desktop, while the pet sprite itself remains interactive. It also supports IME-aware input handling and displays vibe-heart animations and plugin messages mirrored from the main renderer.
apps/desktop/src/app/pet-overlay · high confidence
Introduce live A/B evaluation harness for core tool deferral
Added a new evaluation harness in evals/core\_tool\_deferral to measure the impact of deferring core tools (PR \#97979). The harness runs real in-process agents against baseline and PR code paths, scoring accuracy, API turns, token usage, and wall time across multiple models. It includes a 14-task battery (tasks.py), an isolated worker (worker.py) that sandboxes host-side effects like crontab and HOME writes (sandbox.py), an orchestrator for parallel, resume-safe execution (orchestrator.py), and a reporting tool (report.py). Initial results show reduced token usage with flat overall accuracy, though a regression in the 'clarify' tool's structured interaction was identified.
_evals/core\_tool\deferral · high confidence
Introduce local OpenAI-compatible proxy for OAuth upstreams
Adds a new \hermes proxy\ CLI subcommand and server that acts as a local OpenAI-compatible forwarder, attaching the user's OAuth-authenticated credentials to requests sent to upstream providers (such as Nous or xAI). The proxy listens on localhost by default (port 8645), supports \start\, \status\, and \providers\ subcommands, and enforces strict security by validating Host headers and Origin/Site headers to prevent unauthorized cross-origin browser access.
_hermes\cli/proxy · high confidence
Introduce managed llama.cpp runtime for local models
Adds a new local runtime package that installs, supervises, and manages a llama.cpp server. This enables local model inference with machine-scoped model storage, automatic GPU and memory budgeting, and support for vision capabilities and multi-token prediction (MTP) via a built-in catalog. The runtime handles server lifecycle, detects existing servers, and provides real-time load progress and context growth features.
_hermes\_cli/local\runtime · high confidence
Introduce pet engine core for agent activity visualization
A new \agent.pet\ package provides the shared core logic for displaying animated sprite 'pets' across CLI, TUI, and desktop interfaces. This includes geometry constants for sprite atlases, a state-mapping system that translates agent activity signals (such as errors, tool execution, or completion) into specific animation states, and utilities for managing sprite scaling and row resolution.
agent/pet · high confidence
Introduce pet generation UI components
Added a suite of new React components in the pet-generation flow: DraftGrid for selecting and remixing draft images, EmptyHint for example prompts, GenerateUnavailable for setup guidance, HatchPreview for naming and adopting the final pet, HatchingView for progress tracking, ProviderPicker for selecting image backends, and ReferenceChip for managing reference images.
apps/desktop/src/app/pet-generate/components · high confidence
Introduce project verification subsystem
The agent/verify module now provides a verification subsystem that detects the project's run recipe (e.g., package manager) and persists the result in a user-editable environment manifest at .hermes/environment.json. This allows the system to prioritize a saved manifest over fresh static detection, ensuring consistent verification behavior across runs.
agent/verify · high confidence
Introduce provider-specific transport layer for response normalization
The agent now uses a pluggable transport registry to handle provider-specific message conversion, tool schema mapping, and response normalization. New transport modules for Anthropic Messages, AWS Bedrock Converse, OpenAI Chat Completions, and OpenAI Responses (Codex) centralize dialect translation, ensuring consistent behavior across providers. This change also migrates the internal MCP server to the MCP 2.x SDK and exposes a curated set of Hermes tools (including the new Kanban review lifecycle) to the Codex runtime.
agent/transports · high confidence
Introduce the unified 'pm' package manager
Hermes now uses a new, unified package manager (pm) to manage all dependencies, including Python virtual environments, tool binaries (uv, Python), and plugins. This replaces the previous ad-hoc dependency handling with a centralized system that uses a lockfile (pm/lock.json) for precise version and hash pinning, supports optional default packages with user opt-outs, and provides commands for locking, installing, and repairing the environment.
pm · high confidence
Introduce typed roster source helpers and status logic
Added \roster-source-fetch.ts\ and \roster-source-status.ts\ to the Electron main process to standardize how roster data is fetched and how source reachability is determined. The new \fetchRosterSourceData\ utility concurrently fetches profiles and the install ID, while \rosterSourceStatus\ provides a consistent way to evaluate whether a source is reachable, has an error, or requires sign-in. These helpers support the broader multi-source agent and connection registry work by centralizing source state evaluation.
apps/desktop/electron (part 6 of 6) · high confidence
Introduces a one-time in-app tips system for desktop onboarding
The desktop app now features a guided tips system that introduces key interface elements—such as the sidebar, composer, and right pane—through a single, non-repeating rotation. This system ensures tips are only shown when their target elements are visible and prevents the same tip from appearing again after it has been viewed or dismissed, effectively turning the tips into a one-time introductory tour rather than a recurring notification.
apps/desktop/src/lib/tips · high confidence
Introduces interactive setup cards for onboarding beats
Adds a new set of UI components in the assistant-ui setup-beat module that present interactive cards during the first-run onboarding flow. Users can now enter their name, pick a theme accent and layout, select plugins to install, and choose a tour or build type directly within the chat interface. The system displays these cards as interactive elements (chips, inputs, buttons) and collapses them into a one-line receipt showing the user's choices once the beat is settled, while correctly handling user text input and installation states.
apps/desktop/src/components/assistant-ui/setup-beat · high confidence
Introduces main-process connection caches for the multi-connection registry
The desktop app now maintains a main-process cache keyed by connection ID to track which machine each connection points to. This cache supports the new multi-connection registry by ensuring that lookups remain valid only while a connection ID maps to the same machine, and it is invalidated when a connection is removed or recycled. This change underpins the ability to manage multiple named agent sources and their associated gateways from a single desktop instance.
apps/desktop/electron (part 2 of 6) · high confidence
Introduction of built-in gateway hooks module
A new module for built-in gateway hooks has been added to the codebase, serving as the registration point for hooks that are always active. This change establishes the structural foundation for the hook system, though the specific implementation details of individual hooks are not visible in this excerpt.
_gateway/builtin\hooks · medium confidence
Introduction of hermes\_platform for centralized machine facts and resource resolution
The hermes\_platform module is now available to provide a unified source for host machine facts (such as OS family, native architecture, and CPU identity) and resource resolution. This module ensures that hardware recognition is based on actual system capabilities rather than environment variables, distinguishing the control host from terminal execution targets. It also establishes a strict policy where resource resolvers must be located within hermes\_platform/resolver/ to ensure managed runtime resolution.
_hermes\platform · high confidence
Kanban plugin: connection-scoped data layer, board switcher, and native notifications
The Kanban plugin now scopes all data queries and event streams to the active gateway connection, ensuring that switching connections cleanly isolates caches and board selections without cross-contamination. A new board switcher component allows users to create, rename, import, export, and delete boards, with its placement dynamically adapting to either the workspace page header or a split route tile. Additionally, the plugin now delivers native OS notifications for terminal task events such as completions, blockers, and failures, using an in-memory cursor to deduplicate replayed history and ensure users are alerted only to new, relevant state changes.
apps/desktop/src/plugins/kanban · high confidence
Live theme accent and layout preview cards in the chat
Users can now select and preview theme accents and layout presets directly within the assistant chat interface. The new AccentPickerCard displays eight accent swatches plus a custom color well, allowing users to see the accent applied live to the current theme before committing. The LayoutPickerCard offers two first-run layout options (Basic and Elite), providing a live thumbnail preview of the pane arrangement; selecting Elite specifically opens the terminal and file tree in the preview. Both cards use a 'Continue' button to confirm the choice, ensuring that previews are temporary and do not persist until explicitly committed.
apps/desktop/src/components/assistant-ui/theme-picker · high confidence
Local proxy now supports Nous Portal and xAI Grok upstream adapters
The local proxy server now includes built-in adapters for Nous Portal and xAI Grok, allowing users to forward requests to these providers via \hermes proxy start --provider nous\ or \--provider xai\. The Nous adapter handles JWT-based inference with automatic credential refresh and URL validation, while the xAI adapter manages OAuth credentials from a pooled store, supporting retry logic for 401 and 429 responses. Both adapters restrict upstream traffic to specific OpenAI-compatible paths (e.g., \/chat/completions\, \/models\) and provide clear error messages if authentication is missing.
_hermes\cli/proxy/adapters · high confidence
Messaging page overhaul: profile scoping, status filtering, and Telegram QR setup
The Desktop messaging settings page now explicitly scopes API requests to the active profile, preventing credential leakage when switching profiles. Users can filter the platform list by status (e.g., 'Needs attention', 'Connected') and edit allowed user lists using a dedicated multi-input field. A new Telegram QR quick-setup flow replaces manual bot configuration, and platform icons now use brand-specific glyphs with fallback monograms.
apps/desktop/src/app/messaging · high confidence
Model providers are now pluggable via a standardized plugin system
The model-providers directory has been restructured into a plugin architecture where each provider (e.g., Anthropic, DeepSeek, OpenRouter) is a self-contained subdirectory containing a Python profile and a manifest. This allows the system to automatically discover and register providers, enabling users to add new providers by dropping files into the plugins directory and to override built-in providers with custom implementations.
plugins/model-providers · high confidence
Native HUD modifier gesture detection on macOS, Windows, and X11
The desktop app now uses dedicated native helper processes to detect a clean, passive modifier-tap gesture (e.g., pressing and releasing two specific modifier keys) to summon the HUD. Platform-specific implementations (macOS Cocoa, Windows Raw Input, and X11 XI2) monitor only physical key state without capturing characters or injecting input, enforcing strict rules such as blocking on extra modifiers, held keys, mouse activity, or gestures exceeding 500ms. When the gesture is recognized, the helpers emit a bounded JSON message to the parent process, enabling the HUD to appear without interfering with other applications.
apps/desktop/electron/native · high confidence
New 'Hello Runtime' plugin example with runtime-loaded status bar chip
A new example plugin named 'Hello Runtime' has been added to the desktop application, demonstrating how plugins can be loaded at runtime via raw text blob imports and SDK injection. This plugin registers a status bar chip in the bottom-right corner that displays a dynamic label ('⚡ runtime' or '⚡') and reflects the current gateway connection status. Users can configure the plugin through the Settings \> Plugins interface, where they can toggle the chip's visibility and choose its display style, with all preferences persisted locally.
apps/desktop/src/plugins/hello-runtime · high confidence
New Agents panel for monitoring subagent activity
A new Agents view has been added to the desktop app, providing a dedicated panel to monitor live subagent work and delegation history. This feature displays a hierarchical tree of subagents, showing their status (running, queued, failed, or done) via distinct visual indicators, and includes details such as duration, token usage, and age. The panel intelligently aggregates subagents by session, ensuring that active background work remains visible while preventing stale history from inactive sessions from cluttering the view. It also supports grouping of delegation batches and handles both modern backend tags and legacy heuristics for consistent display.
apps/desktop/src/app/agents · high confidence
New Artifacts page with precise indexing and remote support
A new Artifacts page has been added to the desktop app, providing a centralized view of files, images, and links generated during sessions. The indexing logic has been refined to reduce noise by excluding pip cache files and passive links in tool output, while explicitly capturing files produced by the terminal tool and Office documents. It also correctly handles remote gateway artifacts by downloading them via the gateway rather than opening them externally, and fixes timestamp rendering issues.
apps/desktop/src/app/artifacts · high confidence
New Connectors page with multi-account management and server addition
The Connectors page now supports managing multiple accounts per connector, allowing users to list, rename, and remove accounts directly from the UI. The AccountsSection component handles account status indicators (active, expired, failed) and provides reconnect actions for repairable states. A new AddServerDialog enables users to add custom MCP servers through a guided form or raw JSON editor, with support for both stdio and HTTP transports, authentication options (bearer, oauth), and environment variable configuration. The directory displays connectors with status badges, catalog verification marks, and prefetching capabilities.
apps/desktop/src/app/capabilities/connectors · high confidence
New Core-Toolset A/B Evaluation Harness
A new A/B evaluation harness has been added to the \evals/toolperf\_abeval\ directory to measure the impact of core-toolset changes on model efficiency. This tool runs a battery of nine error-inducing tasks—designed to trigger specific failure classes like venv confusion, ambiguous edits, and hidden-file searches—against both a baseline and a fixes branch. It measures key performance indicators including LLM turns, tool calls, tool errors, and wall-clock time using NeMo Relay ATOF traces, allowing users to verify that changes reduce model waste without introducing regressions.
_evals/toolperf\abeval · high confidence
New Hermes Achievements dashboard plugin
A new bundled dashboard plugin, hermes-achievements, adds a collectible badge system to the Hermes interface. It scans local Hermes session history to unlock tiered achievements (Copper through Olympian) based on agent behavior such as tool usage, debugging patterns, and model interactions. The plugin provides a dedicated dashboard tab with a catalog of 60+ achievements, real-time scan progress indicators, and client-side share cards for unlocked badges.
plugins/hermes-achievements · high confidence
New MCP server management interface with catalog installation and live logs
The desktop app now provides a dedicated MCP management capability that replaces the previous tab. Users can browse and install servers from a bundled catalog, edit the mcp.json configuration via a JSON editor that highlights individual server blocks, and monitor real-time logs for each server. The interface displays server status indicators (ok, error, needs-auth, probing) and shows usage analytics, including token costs and tool call counts over the last 30 days.
apps/desktop/src/app/capabilities/mcp · high confidence
New Nix build infrastructure and declarative modules
This change introduces the core Nix build system for Hermes, including the main agent derivation, a dev shell, and a build-time verification suite. It adds declarative configuration modules for both NixOS and Home Manager, allowing users to install the CLI and desktop app via \programs.hermes-agent\ and manage the agent service via \services.hermes-agent\. The desktop application is now built as an Electron wrapper that bundles the renderer and native dependencies, while the dev shell provides a self-contained environment with pinned tools and a Wayland compositor for testing.
nix · high confidence
New Pokédex-style pet generation overlay
The pet generation interface has been moved from a cramped command-palette popover into a dedicated, full-screen Pokédex-style overlay dialog. This new view provides a more spacious layout for the generate, hatch, and adopt flow, featuring a live-streaming draft grid, animated egg/hatching effects, and a clearer separation of stages (idle, generating, hatching, preview, adopt). The overlay handles background processing gracefully, allowing users to close the dialog while generation continues, and includes specific UI components for reference image handling, provider selection, and error states.
apps/desktop/src/app/pet-generate · high confidence
New Tools catalog page with usage-based sorting and scoped configuration
The Capabilities section now includes a dedicated Tools tab that presents available toolsets as a browsable catalog. Toolsets are organized into a feature row (showing either the four most-used toolsets or a set of essentials) and themed groups (Work, Create, Recall, Connect). Each toolset entry displays a badge with its call count over the last 365 days, which is cached and refreshed on demand. Users can enable or disable individual toolsets or apply bulk changes, with optimistic UI updates that revert on error. Selecting a toolset opens a detail dialog containing specific configuration panels (e.g., for vision, computer use, browser, or terminal) and a general config editor. The catalog respects the current profile scope, ensuring that toolset lists and usage analytics are isolated per profile.
apps/desktop/src/app/capabilities/toolsets · high confidence
New TypeScript type definitions for desktop API contracts
The desktop application now includes a comprehensive set of TypeScript interfaces in \apps/desktop/src/types/hermes.ts\ that define the structure of responses from the backend API. These types cover configuration schemas, audio transcription and TTS lease management, OAuth provider status and flows (including free-tier handoffs), and session-related data. This change provides strict type safety for the desktop client's communication with the backend, ensuring that UI components and logic correctly handle the expected data shapes for features like authentication, voice settings, and configuration management.
apps/desktop/src/types · high confidence
New Windows MSIX update strategy with secure channel support
The desktop updater now includes a new \AppInstallerStrategy\ for Windows MSIX updates, handling the full lifecycle of checking for updates, staging the package, and handing off to the OS App Installer. This strategy enforces strict security by verifying artifact digests and sizes, validating native package signatures against the publisher, and rejecting unsafe or authority-mismatched feed URIs. It also introduces a \ChannelStrategy\ to manage dynamic channel resolution, supporting in-place upgrades for active channels and displaying a notice for discontinued channels. Comprehensive tests cover the update flow, error recovery, and secure download verification.
apps/desktop/electron/updater · high confidence
New assistant-ui components for catalog installs, connectors, and artifacts
The assistant-ui directory now includes new components that bring catalog plugin/skill installation, connector management, and artifact handling into the chat transcript. CatalogInstallTool and CatalogAdvancedDialog render a card for the manage\_catalog tool, allowing users to install plugins and skills with advanced options (profile pinning, force install, credential entry) and showing install status or retry actions. ConnectorTool and ConnectorDirectoryCard handle the manage\_connections tool, displaying a directory of hosted connectors with suggested ordering, sign-in flows, and connection status, while ConnectorOffer surfaces named-account notices and rename links. ArtifactCard replaces fenced code blocks with versioned, openable cards that track streaming state and register artifacts in the session store. Additionally, AnsiText renders ANSI SGR color/bold codes in assistant text, and ChatEmptySlot mounts all empty-transcript contributors to prevent silent ownership suppression.
apps/desktop/src/components/assistant-ui · high confidence
New bundled memory providers: ByteRover and RetainDB
The memory plugin system now includes two new bundled providers. ByteRover provides persistent, hierarchical memory via the local \brv\ CLI, supporting local-first storage with optional cloud sync. RetainDB offers cloud-based long-term memory with hybrid search (vector, BM25, reranking), a durable SQLite write-behind queue, and file management tools. Both providers are automatically discovered alongside existing ones and can be selected via the configuration or dashboard.
plugins/memory · high confidence
New bundled web search providers and keyless fallback ring
The web search system now includes several new bundled providers: Brave Free (search-only), DuckDuckGo via the ddgs package (isolated in a disposable child process to prevent GIL blocking), Exa (search and extract), Firecrawl (search, extract, and crawl with keyless cloud support), Keenable (search and extract), OpenAI Native (server-side web\_search via Codex Responses), and Parallel (async extract). A new keyless fallback ring allows anonymous users to perform searches and extractions using free tiers from Exa, Parallel, Firecrawl, and Keenable when no API keys are configured, with automatic failover on rate limits or refusals. The DuckDuckGo provider now runs in a separate process with a hard wall-clock timeout to avoid freezing the agent loop.
plugins/web · high confidence
New chat UI primitives: activity timers, code cards, and diff rendering
The chat component now includes a suite of new UI primitives to enhance the conversation experience. An activity timer system tracks and displays elapsed time for ongoing operations, persisting state across component remounts to ensure accuracy even when items scroll out of view. Code blocks are now rendered within dedicated 'code cards' that provide a consistent, tinted surface, while a new syntax-highlighted diff viewer allows users to see file changes directly in the chat with a robust fallback to plain colored text if the highlighter fails to load. Additionally, a compact markdown renderer standardizes the display of tool-generated content, ensuring proper typography and overflow handling for tables and code fences.
apps/desktop/src/components/chat · high confidence
New cloud browser providers: Browser Use, Browserbase, and Firecrawl
The browser plugin now supports three new cloud browser backends: Browser Use, Browserbase, and Firecrawl. Users can configure these via environment variables (e.g., BROWSER\_USE\_API\_KEY, BROWSERBASE\_API\_KEY, FIRECRAWL\_API\_KEY) and select the provider using the browser.cloud\_provider setting. Browser Use supports both direct API keys and a managed Nous tool gateway. Browserbase offers features like stealth, proxies, and keep-alive, with automatic fallbacks for paid features. Firecrawl provides a remote browser execution backend. A shared base class (CloudBrowserProvider) standardizes session creation and release across these providers.
plugins/browser · high confidence
New codebase navigability evaluation harness
Added a new benchmark suite under \evals/codebase\_navigability\ to measure the cost for an LLM agent to navigate the codebase. The suite includes \bench.py\ to calculate the token cost of locating and reading symbol definitions, \lookup\_sim.py\ to simulate a skilled agent's grep-and-read workflow, \static\_metrics.py\ to analyze code structure (LOC, complexity, import cycles), and \runtime\_bench.py\ to measure startup times and import overhead. This provides a deterministic, offline way to track how refactors impact developer and agent productivity.
_evals/codebase\navigability · high confidence
New compaction evaluation harness for measuring recall accuracy
A new evaluation harness in \evals/compaction\ allows users to measure the actual recall cost of context compaction policies rather than just token reduction. The runner loads transcripts, generates factual recall questions from the region to be summarized, and tests various policies (including the default \current+recovery\ path, lean modes, and the Jev decision-model arm) by answering those questions using only the compressed context. It provides a scorecard comparing recall accuracy against tokens retained, along with scripts to reconstruct transcripts from session lineage and simulate repeated compaction cycles.
evals/compaction · high confidence
New contribution system for desktop chrome and panes
The desktop app now uses a contribution-driven architecture to manage its user interface. A new \contrib\ module introduces a controller and wiring system that allows core features and plugins to register UI surfaces—such as panes (files, logs, review), statusbar items, and titlebar tools—into a unified layout tree. This system supports multiple layout presets (Default, Basic, Focus, Terminal Deck, Quad) and enables features like deep-link installation of MCP servers with explicit user consent, plugin preview script permissions, and a new Logs pane with intelligent stick-to-bottom behavior.
apps/desktop/src/app/contrib · high confidence
New dashboard authentication framework with secure session handling
The dashboard now uses a new authentication framework that enforces session verification for all requests when bound to a non-loopback host. This includes a server-rendered login page, support for OAuth and password-based login providers, and secure cookie management with prefix hardening (\_\Host-/\\_Secure-) and PKCE state handling. The framework also introduces an audit log for authentication events, body size limits on public auth routes, and proper handling of reverse proxy prefixes (X-Forwarded-Prefix) for correct URL generation and redirect targets.
_hermes\_cli/dashboard\auth · high confidence
New desktop UI components and boot-failure recovery surfaces
The desktop app now includes several new UI components and recovery surfaces. A \Backdrop\ component provides a subtle visual overlay, while a \BillingBanner\ displays in-stack billing notifications with recovery actions. A \BrandMark\ component renders the theme-aware application icon. The \BootFailureOverlay\ and its supporting logic (\boot-failure-cause\, \boot-failure-reauth\) provide a comprehensive recovery interface for local and remote boot failures, classifying errors into plain-language causes and offering in-place gateway settings or re-authentication. The \ConfirmHost\ component implements an imperative \confirm()\ dialog for user approvals. Finally, the \DesktopInstallOverlay\ manages the first-run installation process, including local bootstrap progress and remote setup choices.
apps/desktop/src/components · high confidence
New desktop UI hooks for image handling, scrolling, and performance
The desktop app introduces several new React hooks in the \hooks\ directory to improve image display, user interaction, and rendering performance. \useMediaImage\ now supports a gateway proxy fallback to load images that fail to load directly from external CDNs, and \useImageDownload\ provides a robust mechanism to save images to disk with correct filenames derived from MIME types. \useGrabScroll\ adds a shared 'grab-to-pan' interaction for overflow containers, while \useResizeObserver\ optimizes performance by sharing a single observer instance across consumers to reduce layout thrash. Additional hooks include \useDelayedTrue\ to prevent flashing loading skeletons, \useViewedInterval\ to pause background UI work when the window is unfocused or hidden, and \useThemeEpoch\ to help consumers re-resolve colors after theme changes.
apps/desktop/src/hooks · high confidence
New deterministic pet generation pipeline with atlas-based spritesheet assembly
The pet generation flow now uses a dedicated pipeline that separates prompt-based image generation from deterministic spritesheet assembly. The new \atlas\ module handles the precise layout of generated frames into a standardized 8x9 grid (1536x1872 pixels) using median-box fitting and background removal, ensuring consistent sprite geometry. Prompt generation has been refined to preserve user-requested tones and styles (e.g., pixel art, plush, 3D toy) while enforcing strict layout constraints to prevent frame overlap or misalignment. This change introduces a more reliable and visually consistent pet creation experience by decoupling the creative generation from the technical assembly process.
agent/pet/generate · high confidence
New dev-only performance diagnostics for rendering and store churn
A new \debug/\ module provides two counters accessible via the browser console (\window.\_\_RENDER\COUNTS\\\ and \window.\\_ATOM\CHURN\\_\) to identify wasted re-renders and unnecessary store notifications. These tools are strictly for development; they are aliased to no-ops in production builds to ensure zero impact on shipped code. The render counter uses \bippy\ to track which components re-rendered and why (changed props, state, or context), while the atom churn counter tracks which nanostores published values that were deep-equal to the previous ones, helping developers spot reference-identity issues.
apps/desktop/src/debug · high confidence
New developer documentation and shared platform infrastructure
Added a comprehensive guide (ADDING\_A\_PLATFORM.md) for integrating new messaging platforms via the plugin system or core code, detailing required adapter methods, interactive UX patterns, and configuration hooks. Introduced shared infrastructure modules (\_shared.py, \_http\_client\_limits.py, access\_policy\_mixin.py) to standardize profile-scoped secret reading, YAML-to-environment bridging, HTTP client keepalive tuning, and DM/group access policies across all platform adapters.
gateway/platforms · high confidence
New evaluation scripts for context compaction analysis
Added a suite of Python scripts in \evals/compaction/scripts\ to evaluate and visualize context compaction behavior. \reconstruct\_lineage.py\ rebuilds full transcripts from session databases, while \replay\_lineage.py\ runs these transcripts through the compactor to measure token reduction and message counts. \codex\_arm.py\ provides a head-to-head evaluation against the Codex CLI's auto-compaction. \jev\_cycles.py\ and \jev\_cycles\_report.py\ simulate and report on repeated compaction cycles to identify performance degradation or 'stuck' states. Finally, \build\_html\_report.py\ generates interactive HTML reports comparing compaction results across different runs or configurations.
evals/compaction/scripts · high confidence
New host platform detection and onboarding summary module
The \hermes\_platform.host\ module now provides a centralized \summary()\ function that gathers machine and user account facts for the desktop first-run onboarding. It detects the operating system, CPU, RAM, GPU class, and specific NVIDIA platforms (RTX Spark, DGX Spark, or discrete) using direct hardware sources (IOKit, sysfs, procfs) without environment variables or subprocesses. It also identifies runtime environments like WSL, Docker, and Termux, and retrieves the user's full name and system locale via native OS APIs.
_hermes\platform/host · high confidence
New in-app tips and local runtime update offers
The desktop app now features an in-app tips system that rotates through a catalog of contextual hints, anchored to specific UI elements and paced to appear only during quiet moments (e.g., when not typing or when the window is focused). This system includes a specific offer for local model runtime updates, which checks for available updates and prompts the user to install them via a dedicated action button. The tips are opt-in by default, respect a 30-day tutorial lifetime (after which they are retired), and include telemetry for user interactions.
apps/desktop/src/components/tips · high confidence
New item classification script for urgent content filtering
Added a new \classify\_items\ script within the cron subsystem that allows users to score and filter candidate items (such as emails or feed entries) by urgency. The script accepts a JSON list of items and plain-language importance criteria, calls an auxiliary LLM model to score them, and outputs only those items exceeding a configurable threshold (default 7/10). This enables cron jobs to silently suppress non-urgent intervals, preventing notification spam while ensuring high-priority items are surfaced.
cron/scripts · high confidence
New layout tree renderer with floating panes and drag-and-drop support
The desktop pane shell now uses a new layout tree renderer that supports floating panes, drag-and-drop session reordering, and a layout edit mode. Floating panes are rendered as fixed-position cards that can be dragged by their header, with position and collapse state persisted across sessions. The layout tree supports splitting, merging, and resizing zones via sash drags, with drag-and-drop reordering of tabs and panes within zones. A layout edit mode is available via a floating edit bar that allows users to reset or change layouts. The renderer also includes keep-alive panes that persist their state across layout changes, and narrow viewport overlays for responsive behavior.
apps/desktop/src/components/pane-shell/tree/renderer · high confidence
New memory provider configuration UI with inline autosave and full-configuration modal
The Memory settings page now includes a dedicated configuration interface for memory providers. Users can manage inline settings (such as API keys, base URLs, and workspace IDs) in a compact panel that autosaves changes on blur, with inline error handling and a retry option if the backend fails to load. For providers with additional options, a 'Full config' button opens a modal that groups all fields by category, supports kind-specific inputs (text, number, boolean, JSON, secret), and saves only the fields that have been edited. The UI also integrates an OAuth connection flow for providers that support it, displaying connection status and allowing reconnection.
apps/desktop/src/app/settings/memory · high confidence
New modular dashboard authentication providers and shared framework
The dashboard authentication system has been restructured into a plugin-based architecture with a shared helper module (\_shared.py) and four distinct provider plugins. Users can now authenticate via the existing Nous Portal OAuth (nous), a new generic self-hosted OpenID Connect provider (self\_hosted) supporting public and confidential clients, a new username/password provider (basic) using stateless HMAC-signed sessions, and a service-credential provider (drain) for the gateway drain-control endpoint. The shared framework standardizes PKCE login flows, token exchange, JWT verification with configurable clock-skew leeway, and session rotation via refresh tokens across all providers.
_plugins/dashboard\auth · high confidence
New modular embed detection system for desktop assistant
The desktop assistant now uses a new, modular provider system to detect and render rich embeds from YouTube, Vimeo, Instagram, Pinterest, TikTok, Spotify, Google Maps, OpenStreetMap, and X/Twitter. This change introduces a dedicated detection layer in \apps/desktop/src/components/assistant-ui/embeds/providers\ that parses URLs into structured descriptors, enabling the UI to render these embeds in isolated iframes or specialized components (like the tweet renderer). The system includes comprehensive unit tests to ensure correct URL parsing and embed URL generation for each supported provider.
apps/desktop/src/components/assistant-ui/embeds/providers · high confidence
New monitoring subsystem with OTLP egress and strict redaction
The agent now includes a dedicated monitoring package that exports gateway health and diagnostic events to an OpenTelemetry Collector via OTLP. This system uses a fire-and-forget emitter to ensure monitoring failures never block the gateway, and enforces strict data hygiene by redacting secrets, PII (emails, phones, UUIDs), and limiting resource attributes before export. It also introduces a stable, persisted install ID to help operators distinguish between different gateway instances in telemetry.
agent/monitoring · high confidence
New pluggable image generation backends (DeepInfra, FAL, Krea, Meta, OpenRouter, xAI)
The image generation system now supports multiple external providers via a unified plugin architecture. Users can generate images using DeepInfra (FLUX, Qwen), FAL.ai (Flux, GPT-Image, Nano Banana), Krea (Krea 2/3 with editing and upscaling), Meta Model API (Muse Image), OpenRouter (GPT-Image-2, Gemini, Grok, etc.), and xAI (Grok Imagine). Each provider is registered as a distinct backend with its own configuration, API key requirements, and model catalog, allowing users to select their preferred image generation service.
_plugins/image\gen · high confidence
New pluggable secret source system with Bitwarden, 1Password, and command integrations
The agent now supports fetching environment credentials from external secret managers at startup via a new pluggable SecretSource interface. Bundled sources include Bitwarden Secrets Manager (with lazy CLI installation, encrypted disk caching, and EU/self-hosted server support), 1Password (resolving op:// references with auth caching), and a generic command source (running a user-configured POSIX helper to parse KEY=VALUE blobs). All sources share a unified caching substrate, error classification, and non-blocking startup behavior, allowing third-party managers to register via plugins.
_agent/secret\sources · high confidence
New plugin SDK capabilities and Kanban badge contributions
Plugins can now interact with the desktop host through several new SDK features: they can execute built-in app actions (like toggling the browser) via \ctx.runAction\, access the user's file system within granted folders using \ctx.files\, and display transient messages in the core pet bubble via \ctx.pet.say\. Additionally, plugins can contribute small text badges to Kanban cards through the \kanban.card.badges\ area, and execute JavaScript in the preview browser tab via \ctx.preview\ (subject to capability declaration and per-origin user consent). The \events.ts\ file also introduces a gateway event tap that allows plugins to subscribe to app-wide events with isolated, error-safe listeners.
apps/desktop/src/contrib · high confidence
New plugin SDK surface for desktop integrations
The desktop application now exposes a comprehensive plugin SDK (\apps/desktop/src/sdk\) that allows third-party plugins to interact with the core application. This includes a capabilities bridge for managing skills, toolsets, and profiles; a composer API for reading, writing, and submitting drafts in specific sessions; and read-only state atoms for monitoring focused session activity, connection status, and turn flags. Plugins can also register custom UI areas (panes, status bar, titlebar), contribute Kanban card badges, render media produced by their tools, and extend the application's internationalization support. The SDK enforces strict scoping, ensuring that plugin actions like composer edits or session openings are confined to their intended targets without hijacking the user's current view.
apps/desktop/src/sdk · high confidence
New post-mortem evaluation harness for agent run forensics and live A/B testing
A new evaluation harness has been added to \evals/postmortem\ to analyze and verify fixes for the 1,393-agent run (tracking issue \#102117). The harness provides two main capabilities: a 'forensics' lane that reads a copy of the Hermes \state.db\ and agent logs to recompute observed metrics (cost, cache behavior, delegation timeouts, tool friction, and rework inventory) without requiring model calls, and a 'live A/B' lane that exercises real code paths against local fake providers to verify specific PR fixes (such as scanner bypasses, context caps, and auth stampedes) via side-by-side PASS/FAIL reporting. The package includes specific probes for cache concurrency, batch failure notices, and goal-loop behavior.
evals/postmortem · high confidence
New project sidebar implementation with worktree support and appearance customization
The project sidebar has been rebuilt to render backend-authoritative project trees, introducing a new \BaseBranchPicker\ for creating worktrees from specific branches and a \ProjectAppearancePicker\ that lets users customize project icons and colors. The new layout distinguishes auto-discovered repositories from explicit projects, supports manual drag-to-reorder with pinned sessions, and provides an 'Undo' option when hiding auto-discovered projects. Session visibility is now paginated to handle large projects efficiently, and the sidebar correctly anchors new sessions to the first repository root for path-less multi-folder projects.
apps/desktop/src/app/chat/sidebar/projects · high confidence
New reasoning-replay consumption evaluation harness
A new evaluation harness has been added to \evals/reasoning\_replay\ to measure whether AI providers and models actually consume (tokenize and process) reasoning traces replayed from earlier assistant turns, rather than just accepting them. The tool sends synthetic and verbatim opaque reasoning carriers to various providers (via direct routes, OpenRouter, and custom endpoints) and measures consumption by comparing prompt-token deltas and canary recall against a baseline with no reasoning. It supports multiple reasoning formats (e.g., \reasoning\_content\, \reasoning\, \reasoning\_details\, Gemini \thought\_part\) and positions (in-loop vs. cross-turn), handles API key rotation and budgeting, and outputs a consumption matrix for analysis.
_evals/reasoning\replay · high confidence
New rebindable keyboard shortcuts system with layout-aware resolution and IME safety
The desktop app introduces a centralized, rebindable keyboard shortcut system that normalizes key combinations across platforms (using 'mod' for Cmd/Ctrl) and respects user keyboard layouts (e.g., Dvorak, AZERTY) by resolving keys based on the character produced rather than the physical key. The system includes robust guards to prevent keybinds from firing during IME composition (e.g., Chinese/Japanese input) or when focused on editable targets, unless explicitly allowed (e.g., for reasoning level adjustments). It supports layered chords (e.g., ⌘1-9 switching tabs when a tab strip is visible, otherwise switching profiles) and exposes a unified settings panel for users to view and rebind shortcuts. New actions include session archiving, dictation, and reasoning level control, all shipped unbound by default to avoid hijacking typing.
apps/desktop/src/lib/keybinds · high confidence
New review pane with scope selection and virtualized file tree
The right sidebar now features a dedicated review pane that lets you inspect and manage code changes across three scopes: Uncommitted, Branch, and Last turn. The pane includes a file tree that automatically virtualizes when a folder contains many files (over 60 nodes) to prevent UI freezing, while small trees remain fully animated. You can stage/unstage individual files or all files, revert changes, and view diffs. For uncommitted changes, a 'ship bar' at the bottom allows you to write commit messages, generate them with AI, commit/push, or hand the work to the coding agent. The branch and last-turn scopes are read-only views.
apps/desktop/src/app/right-sidebar/review · high confidence
New right-rail preview components and agent interaction engine
The right-rail now features a comprehensive preview system including a browser bar with navigation controls, an agent interaction engine (preview-act) for driving web pages via real input events, artifact previews with versioning and sandboxed rendering, and annotation capabilities. This adds the infrastructure for agents to interact with preview content directly within the chat interface.
apps/desktop/src/app/chat/right-rail · high confidence
New security-guidance plugin warns on dangerous code patterns
The new security-guidance plugin scans file writes (write\_file, patch, skill\_manage) for 25 known-dangerous patterns—including unsafe deserialization, command injection, XSS sinks, and disabled TLS verification—and appends a warning to the tool result so the model can self-correct. It runs locally with zero LLM tokens, warns by default (file is still written), and can be enabled via allow-list; set SECURITY\_GUIDANCE\_BLOCK=1 to refuse writes entirely, or SECURITY\_GUIDANCE\_DISABLE=1 to disable the plugin. The pattern data is a verbatim fork from Anthropic's claude-plugins-official (Apache-2.0); the plugin glue code is original NousResearch work (MIT).
plugins/security-guidance · high confidence
New settings subpages and plugin extensibility for About and Appearance
The Settings UI now organizes Appearance into distinct subpages (General, Theme, Typography, Window Layout, Chat Display, Pet) and introduces a plugin contribution system that allows extensions to inject controls into the Appearance page via an \AppearanceExtraSlot\. The About page has been refactored to display version details, update status, and channel selection in a structured layout, with a dedicated uninstall subpage. Additionally, auxiliary task configuration rows now support plugin-registered tasks alongside built-in slots, enabling dynamic model routing for custom capabilities.
apps/desktop/src/app/settings · high confidence
New shared UI primitives for the desktop application
The desktop application now includes a comprehensive set of shared UI components in the \apps/desktop/src/components/ui\ directory, standardizing the interface across the app. This introduces primitives such as \Button\ (with a new \loading\ state that swaps the label for a spinner without layout shift), \Badge\, \Alert\, \Checkbox\ (with indeterminate support), \Chip\, \AvatarChip\ (unifying identity displays across platforms), \CardStack\ (for swipeable, animated item stacks), \ConfirmDialog\ (with secondary actions and timer cleanup), \ActionsMenu\ (synchronizing kebab and context menus), \ColorSwatches\, \AccentSwatch\, and \Command\ (a command palette/picker). These components provide a consistent visual language and interaction model for desktop-specific features like profile customization, session management, and connector setup.
apps/desktop/src/components/ui · high confidence
New shared gateway acceptance evaluation harness
A new evaluation suite for the shared gateway has been added to verify acceptance criteria across 47 scenarios. The suite includes a receipt validator (\verify\_receipt.py\) to ensure complete coverage of required cases, a bootstrap readiness probe (\normal\_boot.py\) to confirm the gateway starts correctly on POSIX systems, and A/B test scripts (\search\_readonly\_ab.py\, \session\_readers\_cli\_ab.py\) to validate that read-only operations do not mutate state. These tools allow users to run automated checks against their gateway builds to ensure compliance with the unified gateway plan.
_evals/shared\gateway · high confidence
New shell components for approval mode, context usage, and gateway status
The desktop shell now includes dedicated UI components for managing approval modes, displaying context usage breakdowns, and monitoring gateway connection status. The approval mode menu allows users to switch between manual, smart, and off modes per profile, with proper synchronization to the gateway. The context usage panel provides detailed token consumption metrics, including category breakdowns and compression counts. The gateway menu panel displays connection states, inference readiness, and provides controls for reconnecting or restarting the gateway, along with a live log tail for troubleshooting.
apps/desktop/src/app/shell · high confidence
New skills authoring standards and macOS Apple integration skills
The skills system now enforces hardline authoring standards (SKILL.md frontmatter, description length, platform gating, and section order) via a new \skills/AGENTS.md\ guide and corresponding tests. Additionally, new bundled macOS skills have been added for Apple Notes, Apple Reminders, Find My, and iMessage, providing terminal-based control for these native apps.
skills · high confidence
New unified context-menu system for desktop
The desktop app now uses a centralized context-menu architecture that replaces scattered native and custom menus with a single, consistent right-click experience. This new system correctly handles links (including URL chips), images, and editable text fields, offering appropriate actions like opening in-app or external browsers, copying content, and editing. It also introduces a plugin-based composer context menu, allowing extensions to contribute custom actions specifically within the text input area, while ensuring menus are properly scoped to their containers (like dialogs) and respect user interface states (such as hiding in-app browser options in the HUD).
apps/desktop/src/app/context-menu · high confidence
New video generation backends: DeepInfra, OpenRouter, and xAI
The video generation plugin now supports three additional backend providers. DeepInfra is added as a new backend using its OpenAI-compatible endpoint with live model discovery. OpenRouter is introduced as a backend that fronts every major video model (such as Veo, Kling, and Hailuo) via a unified API with live catalog discovery. xAI (Grok Imagine) is also added, supporting text-to-video, image-to-video, and video edit/extend capabilities via the xAI async videos API.
_plugins/video\gen · high confidence
New web dashboard documentation and ESLint configuration
Added \web/AGENTS.md\ to document the dashboard architecture, including the embedded TUI via xterm.js, the separation between \dashboard\ and \serve\ modes, and coding rules for new surfaces. Added \web/eslint.config.js\ to configure linting for the React/TypeScript codebase, specifically downgrading \react-hooks\ v7 rules to warnings to allow existing patterns while refactoring.
web · high confidence
Pluggable computer-use provider architecture with default cua driver
The computer-use plugin now supports a pluggable provider system, allowing users to select different backends via the \computer\_use.backend\ configuration setting (defaulting to the built-in \cua\ driver). The \cua\ provider, which uses the cua-driver over MCP for background computer use on macOS, Windows, and Linux, is included by default. Users can install additional providers in their user plugins directory (\\~/.hermes/plugins/\), and the system will discover and load them, with bundled providers taking precedence in case of name collisions. If the configured provider is not found or fails to load, a clear error message lists available options.
_plugins/computer\use · high confidence
Redesigned capabilities catalog with card-based browsing and tiered organization
The Capabilities catalog now features a redesigned browsing interface with card-based layouts, artwork display, and tiered organization (Official, Verified, Community). Users can browse skills and plugins through a visual card interface with hover animations, search functionality, and category filtering. The catalog includes curated featured items, star-based sorting, and visual indicators for installed capabilities. Installation is handled through a streamlined button interface with progress indicators, and detailed information is available through an expandable detail view with source links, screenshots, and metadata.
apps/desktop/src/app/capabilities/catalog · high confidence
Redesigned desktop onboarding flow with free-tier support and local model offers
The desktop onboarding experience has been restructured into a modular component system that guides users through provider selection, sign-in, and model confirmation. A new free-tier introduction flow now appears on first launch for eligible users, complete with setup-notice handling for backend failures and retry logic. The provider picker highlights Nous Portal as the recommended option while hiding others behind a disclosure, and supports manual API-key entry with fallback to manual model-name input when endpoint discovery fails. After onboarding completes, a local-model offer dialog appears after a settle window if the machine is eligible, allowing users to set up a managed local runtime. The UI includes a new 'glyph' animation for status transitions and improved error states with actionable recovery options.
apps/desktop/src/components/onboarding · high confidence
Redesigned desktop shell with contribution-driven architecture and unified UI primitives
The desktop app shell has been refactored to use a contribution-driven model, allowing panes, titlebar items, keybinds, and routes to register via a central registry. This change introduces a new floating HUD system for the command palette and session switcher that correctly handles macOS titlebar drag bands, and establishes a consistent master-detail layout for the Capabilities and Messaging pages with drag-resizable columns. Session opening logic is now unified through a single door that intelligently manages in-place focus, tab stacking, and window popping based on keyboard modifiers, while new layout constants enforce responsive sidebar docking and content width constraints.
apps/desktop/src/app · high confidence
Shared TypeScript utilities and contracts for cross-surface consistency
The \apps/shared/src\ module now provides a unified set of TypeScript utilities and wire contracts used by the Desktop, TUI, and web clients. This includes comprehensive ANSI escape sequence handling (\ansi.ts\) to ensure clean text rendering across surfaces, a robust sRGB color and contrast module (\color.ts\) for consistent theme and accessibility logic, and a lightweight fuzzy-search scorer (\fuzzy.ts\) for identical picker ranking behavior. Additionally, the location introduces shared billing types and settlement logic (\billing-types.ts\, \billing-policy.ts\, \charge-settlement.ts\), catalog curation and installation helpers (\catalog-featured.ts\, \catalog-install.ts\), and a generated gateway contract (\gateway-contract.generated.ts\) that standardizes the wire format between the Python gateway and TypeScript clients.
apps/shared/src · high confidence
Skills now support external mounts and display full content with metadata
Users can now add skills from external directories, which are treated as a distinct 'external' provenance tier that remains editable in place. The Skills tab now displays the complete skill content (including frontmatter metadata like name and version) in the detail pane, rather than just a summary, and provides an editor for skills that are locally managed or externally mounted.
apps/desktop/src/app/capabilities/skills · high confidence
Support for 1Password and Bitwarden as browser credential vaults
The agent now integrates 1Password and Bitwarden as login backends alongside the existing local Fernet store. Users can sign in with logins from these managers, which are unlocked per-session via a masked master-password prompt (or service-account token for 1Password) and never persisted in memory or environment variables. The backends expose login metadata, resolve passwords, and retrieve TOTP codes, while correctly handling URI match rules (e.g., skipping Bitwarden items marked 'Never') and service-account vault requirements for 1Password.
_agent/vault\backends · high confidence
Systematized desktop performance harness replaces one-off scripts
The desktop app now uses a unified performance harness in \apps/desktop/scripts/perf\ to measure rendering, interaction, and startup metrics, replacing the previous dozen ad-hoc \measure-\\ and \profile-\\ scripts. This new system provides a consistent way to benchmark scenarios like cold-start, streaming, keystroke latency, and render churn, and automatically gates CI checks against a committed \baseline.json\ to catch regressions. It supports isolated instance launching via \--spawn\ to avoid conflicts with running app instances, allows production renderer measurements with \--prod\, and includes dedicated benchmarks for the gateway attach RPCs and the image-attach pipeline.
apps/desktop/scripts/perf · high confidence
Systematized desktop performance testing with 15 new scenario harnesses
The desktop performance testing infrastructure has been consolidated into a unified harness, replacing 12 previous one-off scripts with a structured registry of 15 distinct scenarios. This new system provides standardized, automated metrics for key user experiences: it measures cold-start and first-token latency, tracks composer keystroke responsiveness, monitors frame pacing and long tasks during multi-tab streaming and render churn, and evaluates visual stability during session loads and switches. Additional scenarios cover idle-cost (main-thread contention while streaming), live-window observation, and right-pane stress (project tree and terminal interactions), enabling consistent regression detection and baseline tracking across the desktop application.
apps/desktop/scripts/perf/scenarios · high confidence
Themed guided tour engine with spotlight and directional motion
The desktop app now includes a guided tour system that highlights and narrates any DOM element on both the main app surface and the preview pane. The tour engine uses a custom CSS theme to replace the default driver.js look with a glass-fill popover, hairline border, and no shadow, matching the app's design tokens. It features a directional motion system where popovers slide in from the side they land on, and a companion spotlight-blur layer that keeps the highlighted element crisp while desaturating the rest of the screen. The system automatically discovers tourable targets by scanning for explicit markers, ARIA labels, and other affordances, prioritizing stable selectors that survive re-renders. Tours can be started, stepped through, and stopped via a clean API, and they integrate with the app's navigation and pane system to ensure elements are visible before being highlighted.
apps/desktop/src/lib/tour · high confidence
Unified overlay primitives and split-layout navigation
The desktop app now uses a shared set of UI primitives to render all non-settings overlays (such as Cron, Profiles, and Agents) with a consistent visual language. A new \Panel\ component provides a standardized card-based overlay with a floating close button and consistent padding, while \OverlaySplitLayout\ introduces a master-detail interface featuring a persistent left navigation rail on wide screens that collapses into a dropdown on narrow viewports. This change also adds breadcrumb headers for deep navigation hierarchies and ensures right-click context menu parity for all list rows within these overlays.
apps/desktop/src/app/overlays · high confidence
User-installed context engines are now discoverable and loadable
Users can now install custom context engines into the $HERMES\HOME/plugins directory, where they will be automatically discovered alongside bundled engines (with bundled engines taking precedence in case of name collisions). The system scans for valid engine directories by checking for specific contract markers in their \\init\\_.py files, allowing third-party or user-created context processing logic to be selected via the config.yaml context.engine setting without requiring explicit plugin enablement entries.
_plugins/context\engine · high confidence
Webhooks management page with profile-scoped state handling
Added a new Webhooks page in the desktop app that allows users to create, view, enable, and delete webhook subscriptions. The implementation ensures that webhook state is strictly scoped to the active backend profile; switching profiles clears any in-progress creation drafts and ignores completion callbacks from the previous profile to prevent state leakage.
apps/desktop/src/app/webhooks · high confidence
Architecture
CLI subcommand parsers extracted into hermes\_cli/subcommands
The CLI argument parsers have been extracted from the monolithic main module into individual files under hermes\_cli/subcommands/. This refactoring organizes the command-line interface by separating parser builders for each subcommand (such as cron, auth, config, and computer-use) into dedicated modules, improving code maintainability and reducing the complexity of the main entry point.
_hermes\cli/subcommands · high confidence
Dashboard web routes extracted into modular APIRouter files
The dashboard's web server routes have been reorganized from a single monolithic module into separate, extracted APIRouter modules (e.g., actions, analytics, audio, chat\_ws, config\_env) within the new \hermes\_cli/web\_routers\ package. This refactoring improves code maintainability and testability by isolating route logic, while preserving the original mount order and behavior through late-binding dependencies.
_hermes\_cli/web\routers · high confidence
Refactored gateway event handling into per-family modules with comprehensive test coverage
The gateway event dispatcher in the Desktop session message stream has been split from a single monolithic file into separate, per-family handler modules (e.g., \deferred-pick\, \desktop-bridge\, \input-requests\, \lifecycle\). This refactoring isolates specific event types—such as deferred model picks, desktop bridge interactions, and input request cancellations—into dedicated files, each accompanied by new unit tests that verify correct scoping, state updates, and edge-case handling.
apps/desktop/src/app/session/hooks/use-message-stream/gateway-event · high confidence
Behavioural changes
1741 commits (262 fixes) modifying contributors/emails
A change to existing behaviour in contributors/emails — 1741 commits (262 fixs), 2153 files.
(repo-wide) · medium confidence · unverified
Browser comment mode groups annotations by page region and carries element identity
When you save multiple comments in the in-app browser, they are now grouped by the page region they belong to (e.g., hero, pricing, FAQ) rather than arriving as a flat list. This allows the agent to process related comments as a single task instead of creating separate tasks for each one. Each comment now includes the element's CSS selector, HTML markup, and computed styles, giving the agent the context needed to locate the exact source code. Additionally, a new cross-window handoff system ensures that comments saved in a popped-out browser window are correctly routed back to the original composer or group room, even if the main window's active pane has changed.
apps/desktop/src/lib/preview-annotate · high confidence
Command Center session deletion now requires confirmation
Deleting a session from the Command Center's Sessions list now triggers a confirmation dialog before the action is executed, preventing accidental permanent deletion. This change aligns the Command Center's delete behavior with other entry points in the application that already used a shared confirmation dialog. The update includes regression tests to verify that the delete action is only triggered after explicit user confirmation and is cancelled when the dialog is dismissed.
apps/desktop/src/app/command-center · high confidence
Desktop app source code and contract tests are now TypeScript
The \apps/desktop/src\ directory has been fully converted to TypeScript, replacing the previous JavaScript implementation. This change introduces a new typed REST client (\hermes.ts\) that enforces profile and connection scoping for all backend requests, ensuring that settings, cron jobs, and session data are correctly routed to the active gateway. The update also includes comprehensive contract tests verifying this scoping behavior, alongside new TypeScript definitions for the Electron main-process IPC bridge.
apps/desktop/src · high confidence
Docker image migration to s6-overlay and support for non-PID-1 runtimes
The Docker image replaces the legacy tini init system with s6-overlay as the new PID 1, introducing a robust entrypoint dispatcher that automatically detects and supports non-PID-1 environments (such as Fly Machines, Kubernetes, or Nomad) by bypassing s6 supervision when necessary. This change includes a new privilege-drop shim for \docker exec\ commands to prevent file ownership conflicts, a hardened bootstrap hook for UID/GID remapping, and a deprecation warning for the old \entrypoint.sh\ script. Users benefit from improved stability in container orchestration platforms and better handling of host-user file permissions without breaking existing workflows.
docker · high confidence
Fix s6 service supervision for unprivileged Hermes user
The container now correctly allows the unprivileged Hermes user (UID 10000) to manage s6 services, which previously failed with permission errors because s6-supervise created control directories owned by root. A new initialization script (015-supervise-perms) explicitly changes ownership and permissions of static service supervise trees and event directories to the Hermes user. Additionally, the profile reconciliation script (02-reconcile-profiles) now ensures the dynamic service directory and s6-svscan control FIFOs are writable by Hermes, enabling runtime service registration and state reconciliation without requiring root privileges.
docker/cont-init.d · high confidence
Improved context meter reliability and workspace management in the status bar
The status bar now provides a more accurate and resilient context usage gauge that correctly handles session changes, post-compression states, and transient backend failures without showing stale or incorrect numbers. Additionally, users can now change the working directory of the focused chat session directly from the status bar workspace menu, with the option hidden when the session is remote or busy to prevent errors.
apps/desktop/src/app/shell/hooks · high confidence
Improved pane management and layout editing in the desktop app
The desktop app now features a refined layout editing experience and more robust pane handling. When entering layout edit mode, hidden panes are revealed in Advanced mode to allow rearrangement, while Simple mode restricts editing to currently visible panes. A new lifecycle system keeps recently used panes active in the background (hot-hidden) while parking older ones to save resources, with special handling to keep critical panes like the terminal alive. Additionally, workspace scoping for Bot Mode ensures that bot chat tabs are correctly captioned with the bot's name and prevents renaming of canonical bot chats, while active pane selection is remembered per workspace owner to restore the user's previous context.
apps/desktop/src/components/pane-shell · high confidence
Improved session startup targeting and memory management for warm transcripts
New sessions created from the sidebar or Home now correctly inherit the profile and workspace context of the project tree they are opened under, preventing accidental attachment to the last active project or incorrect profile fallbacks. Additionally, the application now bounds memory usage for cached session transcripts using a weighted LRU strategy (based on count and byte size) and includes a liveness probe to safely evict orphaned transcripts that would otherwise remain pinned in memory after connection drops.
apps/desktop/src/app/session · high confidence
Isolated error boundaries for plugin render hooks
The desktop app now wraps plugin contribution renders in dedicated React error boundaries (\ContribBoundary\). If a plugin crashes during render, the failure is contained to that specific UI slot (displaying a localized error state or warning chip) rather than crashing the entire application or affecting other plugins. This ensures that unstable plugin code degrades gracefully without disrupting the host app's stability.
apps/desktop/src/contrib/react · high confidence
Keyboard navigation and selection logic for clarify cards
The clarify card components now support full keyboard interaction, allowing users to navigate options with arrow keys and select choices using letter or number shortcuts. The UI distinguishes between single- and multi-select modes, highlights the agent's recommended option, and ensures that keystrokes are only processed by the active card to prevent accidental selections in background sessions.
apps/desktop/src/components/assistant-ui/clarify/core · high confidence
New contributor attribution system replaces legacy author map
The project has replaced the monolithic \AUTHOR\_MAP\ dictionary in \scripts/releases/authors\_legacy.py\ with a new file-per-email directory structure under \contributors/emails/\. This change eliminates merge conflicts caused by multiple simultaneous PRs editing the same legacy file, as each contributor mapping is now a separate file. The \contributors/README.md\ documents the new process for adding mappings, and a \reattributed.tsv\ file records specific commits that were re-attributed to the 'OutThisLife' GitHub account to resolve previous identity issues. The release tooling now merges these directory entries with the legacy data, with directory entries taking precedence.
contributors · high confidence
Non-blocking shared-metrics consent with per-profile scoping
The shared-metrics consent flow now presents as a non-blocking strip in the composer status bar rather than a launch-blocking modal, allowing users to proceed directly to chat. Consent is scoped to the active profile, ensuring each profile maintains its own opt-in state, and the interface offers two explicit choices ('Share' or 'Off') without pre-selection. The consent prompt is delayed until after the first-run onboarding and tour are complete, and users can view detailed collection explanations within the strip.
apps/desktop/src/components/shared-metrics · high confidence
Optional skills are no longer activated by default
Official skills maintained by Nous Research are now shipped in the repository but are not automatically copied to \~/.hermes/skills/ during setup. Users can discover and install them via the Skills Hub using commands like \hermes skills browse\ and \hermes skills install \<identifier\>\. This keeps the default skill set lean while providing curated, tested skills for niche integrations, experimental features, or those with heavy dependencies.
optional-skills · high confidence
Plugins are isolated from core and managed via a central catalog
The plugin system has been restructured to enforce a strict boundary between core and plugins: several previously bundled providers (OpenViking, Hindsight, Mem0, Supermemory, Honcho) and third-party integrations (Home Assistant, Spotify, Singularity, Langfuse, MCP) have been removed from core and moved to the plugin catalog, requiring installation via the new catalog system. A new plugin host isolation mode allows running third-party plugins in separate per-profile processes for stability, while a durable data directory ensures plugin state survives updates. The change also introduces a central catalog as the sole discovery system, adds lifecycle events for auxiliary LLM calls, and provides thread-safe utilities for plugin authors.
plugins · high confidence
Plugins page consolidates unified packages and adds granular access controls
The Plugins page now groups a plugin's agent and desktop halves into a single row, using a package marker to correctly pair halves even when folder names differ, and displays clear indicators when one half is missing from the current profile. Users can now manage specific permissions directly from the plugin row, including revoking folder access granted via \ctx.files\ and revoking origins allowed to run preview scripts. Additionally, the page renders plugin-declared settings as a configurable form, shows server connection status pills, and provides an Update button with a security-scan override dialog for git-installed plugins.
apps/desktop/src/app/capabilities/plugins · high confidence
Redesigned Capabilities page with multi-connection profile scoping
The Capabilities page has been restructured into a modular layout with four distinct tabs: Skills, Plugins, Connectors, and Toolsets. A new global scope selector allows users to configure capabilities for any profile on any registered gateway without switching the entire application context, supporting both single-profile and multi-connection setups. The page now features a unified search interface, consistent spacing, and specific UI elements like tier badges on install dialogs and clickable catalog cards for browsing.
apps/desktop/src/app/capabilities · high confidence
Redesigned bootstrap installer UI with improved progress tracking and error handling
The bootstrap installer now features a redesigned user interface across its key screens (Welcome, Progress, Success, and Failure). The Progress screen provides a detailed stage-by-stage breakdown with live elapsed timers and a collapsible log panel, while the Success screen now properly surfaces errors if launching the desktop app fails, preventing silent failures. The Failure screen offers clear retry and log-opening options, and the Welcome screen presents a consistent brand experience.
apps/bootstrap-installer/src/routes · high confidence
Redesigned tool call rows with session-scoped approvals and live subagent delegation
The assistant UI's tool row rendering has been overhauled to provide clearer, more interactive feedback for complex agent actions. Approval requests for dangerous commands (like terminal or code execution) are now presented in a dedicated, session-scoped stack that keeps the user's focus on the active pane and displays the actual command being run. Delegated tasks (subagents) are no longer hidden behind generic JSON blobs; instead, they render as a live, expandable list of child agents showing their individual goals, status, and real-time activity tickers. The fallback renderer also now correctly handles local vision images, distinguishes terminal exit codes from actual failures, and scopes preview artifacts to the correct session, ensuring a consistent and transparent experience across all tool interactions.
apps/desktop/src/components/assistant-ui/tool · high confidence
Refactored remote setup into modular hooks and UI components
The remote setup logic in the desktop app has been restructured to improve maintainability and state management. The previous monolithic implementation was split into distinct, reusable React hooks: \useRemoteProbe\ handles connection discovery, \useRemoteOAuth\ manages authentication flows (including pre-save sign-ins for registry drafts), and \useRemoteConnectionTest\ validates connectivity. These hooks are now consumed by a new \RemoteSetupFields\ component for form inputs and a \FirstRunRemoteSetup\ component for the initial onboarding flow. This change ensures that stale probe results and login attempts are correctly invalidated when the user changes URLs or switches hosts (first-run, settings, registry), preventing race conditions and ensuring that session cookies are written to the correct storage jar for unsaved registry drafts.
apps/desktop/src/components/remote-setup · high confidence
Refined first-run onboarding flow and layout preview behavior
The first-run onboarding experience now uses a dedicated demo layout that is isolated from the user's saved settings; previewing a layout during the intro no longer permanently alters the user's workspace until they explicitly commit the change. The intro sequence includes a typewriter-style greeting overlay that transitions smoothly into the chat, and the system now correctly handles resumed onboarding runs by skipping the demo layout and returning to the user's previous state. Additionally, switching between Simple and Advanced interface modes during the intro preserves the layout associated with each mode, ensuring that a user's layout choice in one mode is not overwritten by a preview in the other.
apps/desktop/src/components/onboarding/first-run · high confidence
Right sidebar file management and session-aware tree behavior
The right sidebar now features a unified file-actions system that provides a context menu for files and folders, supporting actions like revealing in the file manager, copying paths, renaming, and deleting on the local filesystem, as well as downloading remote files. The file tree is now strictly scoped to the currently focused session tile's working directory, ensuring that files are isolated across connections and that the tree correctly reflects the active workspace even when switching between tiles. Additionally, the sidebar includes a toggle to show or hide gitignored files and retains the focused tile's current working directory when interacting with the sidebar chrome.
apps/desktop/src/app/right-sidebar · high confidence
Stabilized file tree and per-project gitignore visibility
The right-sidebar file tree now uses a single, app-lifetime drag-and-drop manager to prevent crashes when the tree remounts (e.g., on CWD changes). It also introduces a per-project toggle to show or hide gitignored files, with the preference scoped to the active connection so that identical paths on different backends are handled independently. Additionally, the tree now keeps nested repository roots visible even if the parent repository's .gitignore would otherwise hide them, and it correctly handles Windows path casing differences during gitignore filtering.
apps/desktop/src/app/right-sidebar/files · high confidence
Systematized desktop performance harness with regression gates
The desktop performance testing infrastructure has been consolidated into a unified harness located in \apps/desktop/scripts/perf/lib\. This change replaces scattered, one-off measurement scripts with a shared library that includes a centralized Chrome DevTools Protocol (CDP) client, isolated instance launching (supporting both dev and production builds), and statistical analysis utilities. A key addition is the baseline comparison system (\baseline.mjs\), which allows performance metrics to be compared against committed baselines to detect regressions in cold-start, streaming, and typing latency. This provides a more reliable and consistent way to measure and guard against performance drift in the desktop application.
apps/desktop/scripts/perf/lib · high confidence
Unified profile management with safe create, rename, and delete flows
The Manage Profiles page and sidebar rail now share a single set of dialogs for creating, renaming, and deleting profiles, ensuring consistent behavior across entry points. Creating a profile now supports cloning from an existing one and optionally setting a SOUL.md persona, with fallback logic for older backends that do not support initial persona assignment. Renaming profiles is now safe: the system retires old renderer sockets and migrates session tiles before the rename executes, preventing the old profile from being resurrected or leaving tabs pointing to a dead backend. Deleting a profile reliably falls back to the default profile if the deleted one was active, retires local gateways, and drops persisted tiles to prevent bots from recreating the deleted directory on relaunch.
apps/desktop/src/app/profiles · high confidence
Unified search and layout system for Capabilities tabs
The Capabilities interface now uses a single, consistent search field and spacing rhythm across all tabs. A new \CapabilitySearch\ component provides a standardized search input with a scope badge and action area, while \CatalogSurface\ renders catalog sections using a shared card layout. A new CSS file enforces a single content column with consistent vertical rhythm and spacing for all Capabilities tabs, ensuring a uniform visual experience.
apps/desktop/src/app/capabilities/ui · high confidence
macOS installer now launches the desktop app with microphone access and hides its own Dock icon
The macOS bootstrap installer (now named "Hermes") has been updated to function as a launcher for the desktop application. It includes the necessary entitlements and usage descriptions to request microphone access for voice conversations, and it is configured as a background UI element (LSUIElement) so that it does not appear as a separate icon in the macOS Dock when running.
apps/bootstrap-installer/src-tauri · high confidence
Fixes
9 commits (5 fixes) fixing apps/bootstrap-installer/src-tauri/icons
A fix in apps/bootstrap-installer/src-tauri/icons — 9 commits (5 fixs), 5 files.
(repo-wide) · low confidence · unverified
Composer completions route to the session owner to fix cross-profile lookup failures
Slash and @-mention completions in the desktop composer now route to the specific session owner (connection and profile) rather than the window's ambient gateway socket. This fixes an issue where non-primary-profile windows received 'No matches' for all lookups because the ambient socket's authority refused another profile's session\_id. The change ensures that catalog and path completions are served by the correct backend, and it introduces a cache-key mechanism to prevent stale ambient results from being served once the true owner is resolved.
apps/desktop/src/app/chat/composer/hooks · high confidence
Connectors page data layer refactored for multi-account and profile isolation
The data layer for the Connectors page has been restructured to support multiple accounts per connector and strict isolation between user profiles. Account operations (connect, reconnect, rename) now target accounts by their unique ID rather than a mutable name, ensuring that repairs and retries affect the correct account even if it has been renamed. Sign-in operations are scoped to specific profiles, preventing a sign-in in one profile from locking or interfering with accounts in another. The system also introduces a retry mechanism for failed or expired sign-ins, allowing users to retry the exact same request to mint a fresh link for the specific account that failed. Additionally, deep links for account renaming now carry the account ID to prevent ambiguity, and the UI logic ensures that only the page instance matching the link's profile scope can act on it.
apps/desktop/src/app/capabilities/connectors/data · high confidence
Dashboard service lifecycle and security hardening in s6-overlay
The dashboard is now managed as a supervised s6 service with proper lifecycle handling: when the HERMES\_DASHBOARD environment variable is disabled, the service slot remains in a 'down' state rather than restarting in a loop. Security has been hardened by removing support for the HERMES\_DASHBOARD\_INSECURE flag, which previously allowed unauthenticated access on non-loopback interfaces; operators must now configure authentication via basic auth (HERMES\_DASHBOARD\_BASIC\_AUTH\_USERNAME/PASSWORD) or OAuth (HERMES\_DASHBOARD\_OAUTH\_CLIENT\_ID). Additionally, the HOME directory is explicitly set to /opt/data before privilege dropping to ensure state files are stored correctly under the unprivileged hermes user.
docker/s6-rc.d · high confidence
Desktop Electron authentication, notifications, and plugin file access hardening
This release significantly improves the security and reliability of the desktop application by hardening how it handles sensitive data and external interactions. Authentication flows are now more robust, with native OAuth login, token storage, and session recovery refined to prevent credential leaks and handle edge cases like cold-start races. System notifications have been reworked to ensure they are delivered correctly across platforms and properly cleaned up on quit. Additionally, plugin file access is now strictly scoped and guarded, preventing plugins from reading or writing outside their intended directories, while the clipboard and window management logic has been updated to align with modern Electron standards.
apps/desktop/electron (part 4 of 6) · high confidence · unverified
Desktop Electron stability and reliability fixes
This release delivers a large batch of stability and reliability fixes for the Desktop Electron application. Key improvements include better handling of backend lifecycle and connection states, fixes for window management and UI scaling, and enhanced error reporting for SSH and remote gateway connections. The update also addresses various platform-specific issues on Windows, macOS, and Linux, ensuring a smoother and more consistent experience across different operating systems.
apps/desktop/electron (part 1 of 6) · high confidence · unverified
Disable MSIX file and registry virtualization for Hermes desktop app
The Hermes desktop app's MSIX package manifest now explicitly disables file system and registry write virtualization. This ensures the application accesses the user's local application data directly rather than in an isolated virtual view, preventing state separation between the MSIX-installed app and the unpackaged CLI/gateway components.
apps/desktop/assets · high confidence
Fix duplicate gateway events and ensure startup latency is reported once per launch
The desktop gateway now prevents duplicate \message.delta\ events from appearing in the UI when multiple sockets connect to the same backend process, by introducing a deduplication gate that tracks events by replay epoch, session ID, and sequence number within a 30-second window. Additionally, the startup latency metric is now reported exactly once per application launch, using a main-process latch to ensure that renderer reloads or reconnections do not trigger duplicate metric submissions.
apps/desktop/src/app/gateway · high confidence
Fix duplicate user messages in durable history
The system now correctly prevents transport-level echo duplicates from being written to the durable user history. Previously, repeated inputs could result in extra user rows in the conversation history; this change ensures that only the actual user-submitted messages are persisted, aligning the stored history with the intended input ownership logic.
_evals/codex\echo · high confidence
Fix session ownership and routing stability in the desktop app
This update resolves a series of session management issues in the desktop client, ensuring that session ownership, routing, and state persistence are handled correctly across various operations. Key improvements include: preserving session owner information during branching and resumption to prevent orphaned sessions; fixing route-based activation to correctly handle sidebar and page state; ensuring transcript occurrences are maintained across session resume and completion; and stabilizing the model picker and session creation flow to respect profile defaults and active connections. Additionally, the changes address edge cases such as preventing deleted sessions from resurrecting, keeping the active chat aligned during gateway reconnects, and ensuring that background session messages do not bleed into the active transcript. These fixes collectively enhance the reliability and consistency of session handling in the desktop application.
apps/desktop/src/app/session/hooks · high confidence
Improved chat transcript reconciliation and attachment handling
The desktop chat now more reliably preserves local assistant errors and system notices across session refreshes, ensuring that failed turns and client-side notifications remain visible even when the backend transcript is rehydrated. Attachment handling has been refined to tolerate backend rewrites of pasted images, preventing duplicate folds and ensuring that captioned screenshots are correctly matched to their replies. Additionally, MEDIA tag parsing now correctly handles file paths with interior spaces and avoids generating dead links from degenerate captures, while reasoning details are extracted more accurately from complex provider envelopes.
apps/desktop/src/lib/chat-messages · high confidence
Improved gateway connection stability and developer experience
The desktop app now maintains the primary gateway socket across Vite Hot Module Replacement (HMR) updates, preventing agent sessions from dropping during development. Connection boot and recovery logic has been hardened to retry initial WebSocket dials against transient transport failures, bound all asynchronous IPC round-trips to prevent UI freezing, and enforce a time-based escalation for reconnect warnings to avoid locking users out during brief network blips.
apps/desktop/src/app/gateway/hooks · high confidence
Improved reliability and clarity in the desktop chat transcript
The desktop chat experience is more stable and easier to follow. Agent initialization failures now display a clear error message in the transcript and release the composer, preventing the user's message from silently vanishing. Duplicate or stale assistant replies are automatically collapsed so the conversation history remains clean. Background and 'between-turn' task results are correctly appended to the originating chat session. Clarification prompts are reliably rendered even after connection interruptions, and compaction progress is accurately tracked and announced. Streaming text is now flushed using a timer instead of animation frames, ensuring consistent updates even when the window is not focused.
apps/desktop/src/app/session/hooks/use-message-stream · high confidence
Improved session status tracking and transcript refresh reliability
The desktop app now more accurately tracks live session status and handles transcript refreshes. New logic ensures that sessions which finish while the app is inactive are correctly reaped, preventing infinite spinners and ensuring the 'your turn' indicator appears promptly. Transcript refreshes are now more robust, preserving completed replies and tool-call results even when the backend returns partial or missing data during a reconnect or background sync. Additionally, the remembered session restore logic has been refined to correctly handle delegate subagent sessions, ensuring they are replaced by their parent sessions or cleared if orphaned, rather than attempting to restore invisible or stale child sessions.
apps/desktop/src/app/contrib/hooks · high confidence
New build-time guards and post-pack hooks for the desktop app
The desktop build process now includes several new scripts in \apps/desktop/scripts\ to prevent shipping broken or misconfigured bundles. \assert-root-install.mjs\ verifies that all critical dependencies (like \vite\, \katex\, \electron\) and matching \react\/\react-dom\ versions are installed before building, avoiding confusing runtime errors from partial installs. \assert-dist-built.mjs\ ensures the renderer output is complete and syntactically valid, checking for duplicate React context providers and invalid ES module chunks. Post-pack hooks (\after-extract.mjs\, \after-pack.mjs\) handle platform-specific finalization: stamping Windows executable identity before ASAR integrity rewriting, restoring macOS locale markers and icons, and sanitizing batch-signing Windows PE certificates. Additionally, \audit-bundle-arch.mjs\ scans native binaries to ensure they match the target architecture, preventing silent emulation failures on user machines.
apps/desktop/scripts · high confidence
Prevent duplicate token refreshes and event-loop blocking in dashboard authentication
This change introduces a single-flight mechanism for token refresh operations, ensuring that concurrent requests using the same stale refresh token trigger only one actual refresh call to the identity provider. This prevents race conditions where multiple tokens might be rotated simultaneously and avoids blocking the event loop during slow provider responses, keeping endpoints like /api/status responsive while a refresh is in progress.
_evals/dashboard\auth · high confidence
Refactored tool row formatting into modular fallback-model components
The tool row rendering logic in the desktop assistant UI has been reorganized from a single large module into a structured folder (\fallback-model\) containing dedicated files for formatting, target detection, and type definitions. This change introduces specific improvements to how tool results are displayed: \browser\_exec\ steps now use a leading comment as their title to match CLI/TUI behavior, and tool output is clamped to 20,000 characters during rendering to prevent UI freezes and out-of-memory errors (while preserving full output for the Copy button). Additionally, \skill\_manage\ refusal errors are now treated as neutral rather than failures, and Windows absolute paths are properly tagged and previewed.
apps/desktop/src/components/assistant-ui/tool/fallback-model · high confidence
Stabilizes desktop chat session management and history navigation
This update introduces robust session tab closing logic (⌘W) that correctly promotes stacked sessions or resets to a fresh draft, preventing dead-end states. It adds a bounded history window feature allowing users to jump to specific messages in long transcripts without replacing the live store, while keeping the edit composer functional during these jumps. Additionally, it fixes the browser pop-out handoff to correctly restore persisted tabs from storage without clobbering them, and ensures the chat composer remains mounted and stable across transient session loaders and profile swaps.
apps/desktop/src/app/chat · high confidence
Stabilizes primary backend routing after profile preference changes
The desktop app now correctly routes requests to the primary backend even after the user changes their default profile preference. Previously, changing the active profile could cause the app to spawn a second backend for the same home directory, leading to resource contention and routing errors. This change introduces a \PrimaryProfilePin\ mechanism that locks the primary backend to the profile it was booted with, ensuring stable routing and preventing duplicate backend spawns.
apps/desktop/electron (part 5 of 6) · high confidence
Test coverage
23878 commits adding/updating tests in tests; Add desktop MCP OAuth integration tests; Add offline integration probe for gateway failure-writer ownership; Add schema footprint evaluation for delegation grouping; Add session\_search schema A/B eval harness and reference results; Added A/B evaluation harness for OpenRouter PKCE authentication; Added A/B evaluation harness for the fixed-prefix re-diet; Added A/B evaluation script for Standard Webhooks signature validation; Added A/B evaluation script for gateway status rendering; Added a development-only demo for cycling credit-notice states; Added benchmark for hosted-room attachment transfer performance; Added desktop bug campaign evaluation probes; Added evaluation harness for Bot Mode native delivery and nested ownership; Added evaluation probe for gateway completion delivery lifecycle; Added evaluation scripts for session persistence and cache eviction controls; Added fixture to reproduce import-latency timing issues; Added live A/B evaluation harnesses for update pipeline recovery scenarios; Added live fixture for testing portal session flows; Added live stress harness for subagent process handoff evaluation; Added liveness evaluation for WebSocket orphan reconnection; Added local evaluation harness for reasoning-shape consistency; Added native tray lifecycle probe fixture; Added property guidance evaluation script; Added shared test fixtures and utilities for the desktop app; Added static analysis test to enforce no native title attributes on buttons; Added streaming update evaluation harness; Added test coverage for dashboard authentication plugins; Added test coverage for the new image generation plugin providers; Added test coverage for video generation plugins; Added test fixtures for Python backend and system CA validation; Added tests for Nix agent integrity and desktop backend selection; Added tests for canonical API client behavior; Added tests for memory plugin configuration, discovery, and profile isolation; Added tests for provider fallback and credential handling; Added tests for the browser provider plugin migration; Added tests for web search plugin migration and registry resolution; Added unit tests for model provider reasoning-effort wiring; Composer componentization and test coverage; Desktop session hooks: comprehensive test coverage for prompt submission and recovery; Expanded test coverage for platform plugin configuration and reliability; Native Electron gateway ticket evaluation harness added; Native e2e probes for Bot Chat delivery and cron root resolution; New A/B evaluation probe for native compaction checkpoint preflight false-positives; New Browser Use A/B benchmark harness; New contract tests for desktop build, install, and macOS entitlements; New deterministic core E2E test suite for Desktop; New end-to-end tests for the Desktop update and first-run experience; New eval harnesses for ACP, API delegation, and approval controls; New hooks and tests for config scoping, keybinds, and profile switching; New library tests and utilities for ANSI parsing, artifact detection, and budgeted animation loops; New live test harness for Tool Search feature evaluation; New offline wire-probe tests for provider configuration, video input handling, and MOA request sanitization; New read-file evaluation harness for hostile workspaces; New tests for agent-avatar cache, inter-agent delivery, and assistant message rendering; New token-accounting evaluation harnesses for image cost calibration and display provenance; Playwright E2E suite for the desktop app; Refactored session actions into modular components with comprehensive test coverage; Sidebar tests cover session archiving, connection switching, and cron job behavior; Tests for the new layout-tree pane management system.
Dependencies
Introduce Tauri-based bootstrap installer and standardize TypeScript tooling across workspaces
The project now includes a new \apps/bootstrap-installer\ workspace, a Tauri 2.x application (Rust backend, React frontend) that provides a native-signed installer UI for the setup process. This addition is accompanied by a comprehensive standardization of the development toolchain: the root \package.json\ and all workspace manifests now pin TypeScript to version 6.0.3, Vite to 8.2.4, and Vitest to 4.1.12. The root manifest also introduces \allowScripts\ configurations to explicitly control native script execution for specific dependencies like \esbuild\ and \node-pty\, and adds \lefthook\ for git hook management. Additionally, a new \tests-js\ workspace has been added to centralize JavaScript/TypeScript testing with Vitest, and the \pm\ Python package has been updated to require Python 3.14.
(dependencies) · high confidence
Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
How this codebase got here
Baseline
- First survey — no prior run to compare against. CAI 47.
Lenses
- Code Health 53
- Architecture 67
- Maturity 76
- Readiness 40
- Security 72
- Domain Modelling 57
- Event Sourcing 100
- Accessibility 44
- Performance 60
Changes since last survey
- 300 commits — 160 feature/other, 140 fixes
By area
- apps/desktop — 25 commits
- contributors/emails — 17 commits
- tests/tools — 13 commits
- plugin-catalog/skill-lineage.yaml — 11 commits
- tests/fixtures — 10 commits
- tests/gateway — 7 commits
- tests/pm — 7 commits
- agent/verification_evidence.py — 6 commits
- gateway/session_cold_review.py — 6 commits
- hermes_cli/model_catalog.py — 6 commits
- plugin-catalog/hermes-jev-routing.yaml — 6 commits
- agent/context_compressor.py — 5 commits
- plugin-catalog/tavily.yaml — 5 commits
- tests/agent — 5 commits
- tests/hermes_cli — 5 commits
- (repo) — 4 commits
- plugin-catalog/gadget.yaml — 4 commits
- plugin-catalog/s2docs.yaml — 4 commits
- website/docs — 4 commits
- (root) — 3 commits
Notable commits
- fix: Merge origin/main into fix/nix-extrapythonpkgs-extend
- fix: Merge pull request #120756 from Lokee86/fix/120629-launchd-force-retry
- fix: Merge pull request #129837 from cl0vrfi3ld/fix/nix-extrapythonpkgs-extend
- fix: catalog: bump live-time to 1.0.1 (timezone resolution fix)
- fix: catalog: limbic 0.9.1 @ e7a3cd3 — memory-quality fixes + version re-track
- fix: catalog: skill-lineage 0.1.4 (6bd396e) - validator clean, three audit fixes
- fix: chore(catalog): re-pin git-review to v0.1.27 (dark theme fixes)
- fix: fix(agent): attribute verify runs behind timeout durations and valued wrapper flags
- fix: fix(agent): ctx.side_question reports through pre/post_auxiliary_call
- fix: fix(agent): drop the unscored-run warning once newer evidence is recorded
- fix: fix(agent): keep asking to verify edits whose directory moved
- fix: fix(agent): keep shell exit codes when recording pipelines
- fix: fix(agent): record piped verification runs and stop double nudges
- fix: fix(agent): redact Context7 API keys by their vendor prefix
- fix: fix(agent): show the run-it-directly advice only when a check run went unscored
- fix: fix(agent): side-question forks release clients only, never close the shared session
- fix: fix(agent): side-question review findings (runtime pin, extra_body tools, hook bracket, lease cap)
- fix: fix(agent): side-question review fixes (unattended turns, profile-bound leases, billing, attribution)
- fix: fix(agent): skip interpreter flags only for Python when matching verify commands
- fix: fix(agent): strip only the wrapper options that still run the child command
- …and 280 more
Architecture
- 0 containers · 1 bounded contexts · 0 dependency edges (baseline)
Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
Survey your own repository
NousResearch/hermes-agent was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.
About this page
- The score is its most recent published measurement, taken on 25 October 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
- Measured at commit 4ea8d23cc9f527332ccb1d55359b27741b6ea51b — the exact code this score is about.
- Scored under rubric-2026.10.5 — the same rubric and the same method as every other entry in this index.
- Measured by watchdog.canine.dev using codehealth-analyzer measure-2026.10.6+preprod-d36780c2d857.