hermes
Download

hermes / changelog

Changelog

RSS feed

Release notes for every version of Hermes IDE, newest first.

  1. v2.0.0

    October 2, 2026latest

    New

    • Hermes 2.0 is a vendor-neutral console for every coding agent, terminal first. Claude Code, Codex, Antigravity, GitHub Copilot, OpenCode, goose, Aider, Kiro, Hermes Agent or any command you add run in their own terminal interface, exactly as they run anywhere else, and Hermes supervises them all the same way: it starts them, tells you which one needs you, records what each turn changed and helps you review and land the result. Hermes follows what your agents report and never answers a prompt on their behalf. The Agent view for Claude stays available as an option. Hermes IDE is not affiliated with Nous Research or its Hermes Agent.
    • Hermes knows today's coding agents and how to install and sign in to each one, now including Antigravity CLI, GitHub Copilot CLI, OpenCode, goose and Hermes Agent.
    • Add any other command-line agent as a Custom agent: it runs in a terminal and shows under its own name in the sidebar.
    • Add more than one account for an agent: each signs in on its own, shows its own models and never borrows another account's sign-in.
    • Every agent starts from one short line that is the same on every shell and OS, instead of its own launch command typed with shell-specific quoting.
    • Quit and reopen Hermes and each agent continues its own conversation in its own folder; if an agent cannot resume, it says so in one line and starts fresh.
    • When an agent refuses to start (a model your account does not have, or you are signed out), the session says why in the agent's own words and offers to retry, pick another model or sign in.
    • An agent stuck at a "do you trust this folder?" prompt is marked as waiting at a startup prompt instead of stalling unseen.
    • Hermes knows when each terminal agent needs approval, asks a question, finishes or fails, with nothing for you to set up.
    • Every session shows one clear status (working, needs approval, asked you, done, error, idle) as a word and a symbol, dimmed when Hermes is only guessing.
    • A badge in the title bar counts the agents blocked on you and opens an inbox of everything waiting, oldest first, where you can peek at each request before jumping to it.
    • Press ⌘I to jump straight to the agent that has waited longest, and press it again to visit the next one. On Windows and Linux that is Ctrl+Shift+I, and Ctrl+Shift+A opens the inbox.
    • Notifications arrive only when an agent is blocked on you or done, never for the session you are looking at, and your computer stays awake while an agent is working.
    • Away from your desk, Hermes can send a short message to a web address you set up (for example ntfy or Telegram) with only the agent, the task and its state, never your prompts or code, including for the session you left in view once you have been away a while.
    • The session header shows which model and permission mode each agent is using.
    • Press ⌘N, describe the task and press Enter: an agent starts in a terminal on its own new branch with your task as its first prompt.
    • The launcher keeps what you typed through a sign-in or a visit to Settings, remembers your usual agent and options for each project, and gives a repeated task the next free branch name instead of an error.
    • A new task's dependencies are ready in seconds when they match your main checkout, and parallel dev servers get their own ports.
    • A .hermes/worktree.toml file in your repository lists setup commands, files to copy such as .env, and a default check, and each new task runs it with a visible log.
    • First launch takes three steps: see which agents are installed and signed in, pick a repository, launch a first task.
    • Settings → Agents shows the same agent check at any time, with install and Sign in actions.
    • Every agent turn is saved in git, including changes made through shell commands, without touching your branch, staged changes or stash, and your own edits between turns are never charged to an agent.
    • A bar under the terminal shows what the last turn changed, with Diff and Restore; Restore shows a preview, says what it sets aside and offers Undo.
    • The Review Desk (⌘G, Ctrl+Shift+G on Windows and Linux) shows everything an agent changed, by turn or by file, and sends your line comments back to the agent that wrote those lines.
    • Revert a single turn after a preview, and see flags on risky changes such as lockfiles, CI workflows, secrets, new binaries and edits to an agent's own configuration.
    • The Land sheet commits on the branch, opens a pull request or merges locally in one step, drafts the commit message from the task and the turns, and can undo afterwards. It lands into the branch the task started from and never over an edit you have not committed.
    • Pull, Abort Merge and Discard keep your uncommitted work, and closing a task never deletes commits or unmerged branches without asking.
    • Hermes's own commits run your repository's hooks, and a refusal is shown exactly as the hook said it.
    • Your checks decide when an agent is done: the commands you list run at the end of each turn and before landing, a failing check shows on the session, and Claude is sent back to fix it, up to three times.
    • Sessions whose recent turns touched the same files show an overlap badge, so you know before merging.
    • Bigger work can follow guided phases (questions, research, design, structure, plan, implement) kept as short markdown files in your repository that you review in your own editor.
    • When a phase waits for your approval it appears in the inbox, and approving it moves the feature to the next phase. Only you can approve or skip a phase, and skipping asks first.
    • Any agent can drive a track from its own terminal, and quick tasks create no extra files.
    • On macOS and Linux, agents keep running while Hermes quits, updates or crashes, and you reattach to them exactly where they were, with their status reports still flowing.
    • Quitting while an agent is working, or a command is still running, asks whether to keep it running or stop it; with only queued tasks it says they are kept.
    • If Hermes's background service stops, your sessions stay listed with their output and come back when you press Restart.

    Improved

    • Each agent session shows which instruction files that agent loads, and one action links CLAUDE.md to AGENTS.md so every agent follows the same project rules.
    • A read-only view lists the MCP servers each agent sees in a project.
    • A context gauge shows how full each agent's context window is, for agents that report it.
    • When an agent hits its usage limit, the session says so with the reset time, and you can continue the task in another agent.
    • Spend is shown exactly where the agent reports it and, where it does not, as an estimate marked "(estimated)" or as "n/a"; an optional spending cap per session or feature acts only on reported spend.
    • Line up tasks with a limit on how many agents run at once; the next one starts when a slot frees up, and queued tasks survive a quit.
    • Many open terminals stay smooth, each agent's memory use is shown on its row, and a layout tiles the agents that are working.
    • Windows and Linux have session and pane shortcuts, Ctrl+Shift+C and V to copy and paste, and right-click Paste in the terminal, and menus show the keys that work on your system.
    • A very large paste arrives whole, and the session follows the folder you cd into in zsh and bash.
    • Dialogs keep the keyboard while they are open, and every new screen is translated and works with only a keyboard or with a screen reader.
    • Settings and the agent check say only what is true, including when an account refuses its default model.
    • Install on Windows with winget install.
    • The Linux AppImage is back, and the macOS and Linux installers are smaller.
    • Typing in a terminal stays responsive while Hermes checks which program is running, most noticeably on Windows.
    • In the Agent view, when Claude stops the view says why (could not start, signed out, exited, busy) and offers Retry or Sign in, and a stuck agent can be force-stopped.
    • Each new task gets its own hermes/<name> branch, and a branch already in use elsewhere asks whether to reuse it, pick a new name or cancel, so two agents never share a checkout by accident.
    • Closing a session with uncommitted changes offers to commit them to the session's branch or archive them, and never stashes your work.
    • A restored session keeps its branch and its folder.
    • Hermes refuses to create a new task when less than 10 GB of disk is free and tells you why.
    • The worktree overview shows how much disk each worktree uses, removes build output on request and clears leftover worktree folders in one action.
    • Starting and tracking agents never changes your agents' own global settings.
    • Every new agent session starts with one safety default (write inside the project, ask before network access or other folders), and an agent running looser than that shows a warning.
    • A session you close just before quitting stays closed, even if Hermes is force-quit right after.

    Removed

    • The four-step New Session wizard is no longer the default way to start: the ⌘N task launcher replaces it, and the full wizard (⌘⇧N) stays available for SSH, tmux and existing branches.
    • The four-step welcome wizard and the "awaiting first signal" screen are replaced by three-step onboarding; theme and privacy choices live in Settings.
    • Stash & Close is gone: closing a session commits its changes to the session's branch or archives them.
    • Gemini CLI is marked legacy; Antigravity CLI is now the main Google agent.
    • The "set up notifications" link and hand-copied setup snippets are gone: agents report their status automatically.
    • The "Task completed" notification and the needs-input indicator in the status bar are replaced by the attention inbox and its notifications.
    • The separate Git panels are replaced by the Review Desk, which keeps the log, stash and conflict views.
    • Rewinding Claude in the Agent view is replaced by restoring a recorded turn, which also covers changes made through shell commands.
    • The context budget meter is replaced by the real context gauge.
    • Hermes no longer edits your global Claude MCP settings; servers are added to the project's own MCP file.
    • --- Published to the beta channel first; promoted to stable after the soak period. Full changelog: https://hermes-ide.com/changelog

  2. v1.4.1

    September 28, 2026

    New

    • Every agent now opens in its own terminal, updates wait until your agents are done, a crash stays inside its pane, and Hermes is private by default.
    • New sessions open every agent, Claude included, in the agent's own terminal interface, exactly as it runs anywhere else.
    • Claude's Agent view is now an option: tick Agent view for Claude on the agent step of New Session to have Hermes render the conversation instead.
    • New Session starts straight on "What do you want to run?"; the separate mode step is gone, and Connect over SSH is a link on that step.
    • Hermes remembers the agent you picked and how you ran it, and preselects both next time.
    • Saved Agent-view sessions still restore in the Agent view.
    • An update no longer relaunches Hermes while an agent is mid-turn or a command is still running; the dialog says how many sessions it is waiting for and offers Relaunch now if you want to install anyway.
    • New builds go to a beta channel first and reach stable after a soak period; pick your channel in Settings → General → Update channel.
    • Before your data is updated, Hermes saves a copy of it and keeps the three newest copies.
    • An older Hermes no longer opens data saved by a newer one: it explains why, offers only Quit, and leaves your data exactly as it was, even after a crash.
    • On Linux, in-app updates now install the .deb package.
    • If something inside one pane fails, only that pane shows an error card with Reload pane and Close pane; the other panes, the session list and the running shells keep working, and Reload brings the pane back with its scrollback.
    • A failure inside a single agent message shows an error card in that message instead of blanking the conversation.
    • If the whole window's interface fails, it can be reloaded without losing any session.
    • Usage analytics is off on a fresh install; nothing is sent unless you opt in, and until then the analytics component is not even loaded.
    • Turning analytics on or off in Settings → Privacy takes effect at once, without a restart.
    • If you do opt in, the three usage events (app started, session created, feature used) are now actually sent; before this release none left the app.
    • New Delete Session Data… action in a session's right-click menu clears what Hermes keeps about that session: saved terminal output, command history, token usage, pinned and remembered context, and agent state. It asks first, and leaves the session and its repository alone.
    • A plugin can no longer read another plugin's stored data, use another plugin's permissions, grant itself permissions, or see what else is installed; installed plugins keep working unchanged.
    • Ctrl+A, Ctrl+D, Ctrl+E, Ctrl+W and the other Ctrl+letter keys now reach the shell or agent in a focused terminal instead of running app shortcuts; Ctrl+D ends input, it no longer splits the pane.
    • App shortcuts on Windows and Linux are now Ctrl+Shift+letter, like Windows Terminal: Ctrl+Shift+D splits, Ctrl+Shift+W closes, Ctrl+Shift+K opens the command palette. The old Ctrl+letter shortcuts still work when no terminal has focus.
    • Shortcuts follow the letters printed on your keyboard (AZERTY, Dvorak, …); the menus and the Keyboard Shortcuts panel show the right keys for your platform.
    • The Keyboard Shortcuts panel and Settings → Shortcuts now list every shortcut the app really has, in your interface language, and nothing that doesn't exist; a shortcuts reference page is included in the docs.
    • MacOS is unchanged.
    • An approval prompt no longer grabs keyboard focus: if the agent asks for permission while you are typing, Enter sends your message and never approves the command by accident.
    • Always allow now saves the rule to the current project's own Claude settings instead of your global ones, so approving a command in one project no longer approves it everywhere; the saved rules are listed under Permissions in the Context tab.
    • Clicking Retry twice now restarts the agent once instead of starting two agent processes for one session.
    • The Manual / Assisted / Auto switch, the Autonomous settings tab and the countdown that ran a predicted command are gone; nothing is ever typed into your terminal on your behalf.
    • Inline suggestions and ghost text no longer appear over an agent CLI (Claude Code, Codex, Gemini CLI or any program the shell started); at a plain shell prompt they still work.
    • New Hermes inline suggestions toggle in Settings → General turns Hermes suggestions off entirely and leaves your shell's own autosuggestions in charge.
    • Plugin side panels and the Git and Files views have their own width (resizable between 240 and 600 px) and a clear edge against the main area; in narrow windows the main area always keeps room to work.
    • Editor font size shortcuts: Cmd/Ctrl+= grows, Cmd/Ctrl+- shrinks, Cmd/Ctrl+0 resets; the size is remembered.
    • GitHub Copilot now means the current Copilot CLI, and its quick actions are the CLI's own slash commands. Having only the retired gh copilot extension no longer counts as Copilot being installed; the card shows how to install the CLI.
    • SSH sessions now honour the jump host you configured, and leaving the SSH user blank lets your ~/.ssh/config decide, as ssh myalias would.
    • Agent-view sessions load CLAUDE.md from every folder attached to the session.
    • Suggestions for starting a VS Code tunnel from the terminal.
    • Hermes-ide --self-test=report.json checks an install (database, bundled runtime, interface, a real shell) and writes a report you can send in; it contains no user or host name.
    • The About dialog says Hermes IDE is not affiliated with Nous Research or its Hermes Agent.

    Fixed

    • Each Claude session follows its own project's transcript; two sessions in two projects no longer both show whichever project was used last.
    • The Workbench layout and session notes are back after a restart.
    • Closing a session that shares a checkout with another session no longer deletes that checkout and its uncommitted work.
    • The branch picker really greys out branches other sessions have checked out, and Next on the folder step waits until Hermes knows whether the folder is a git repository, so a quick click can no longer skip the branch step.
    • : intent commands (for example :status) run again; they had been typing the raw text into the shell.
    • Shell history and shell-type detection reach the suggestions again, so Hermes no longer draws ghost text on top of your shell's own autosuggestions.
    • Splitting a pane that is not the focused one no longer shows the new session in two panes.
    • Unsent image attachments in the agent composer survive switching to a terminal session and back.
    • The working folder in the header updates even when the shell reports it in two pieces, and a stray unfinished escape sequence no longer makes Hermes hold long output in memory.
    • The context-file variable in the launch line uses the session shell's own syntax, so it works in PowerShell and cmd.
    • Session labels, saved hosts and the file explorer no longer show a fake ssh@host or a stray @ when no SSH user is set.
    • Leftover shell-setup files that older versions left in the system temp folder are cleaned up after an upgrade.
    • --- Published to the beta channel first; promoted to stable after the soak period. Full changelog: https://hermes-ide.com/changelog

    Improved

    • Settings, the plugin manager, the code editor, the Agent view, the New Session wizard, side panels and the welcome screens load the first time they are needed, and only the language in use is loaded.
    • Terminal-only use starts no extra agent process at launch.
    • Hermes no longer slows down after days of use: the per-session activity history is capped instead of growing without bound.

  3. v1.4.0

    September 27, 2026

    New

    • Hermes now speaks your language. The interface ships translated into eight languages, and switching takes two clicks — no restart.
    • Open Plugins → Hermes Language Pack and pick your language from the selector: Russian, Spanish, French, German, Portuguese (Brazil), Simplified Chinese, Japanese, or Hindi — alongside English.
    • The switch is instant: every translated surface re-renders in place. Your choice is remembered across launches.
    • Translations cover the surfaces you actually live in: the start screen, the command palette and every shortcut description, the settings tabs end to end, the session creation flow including the SSH and tmux steps, the usage and plan limits panel, the plugin manager, and the prompt composer with its roles, styles, and templates.
    • The language pack is a built-in plugin, so it appears in your installed plugins list where you can see its version and manage it like anything else — but there is nothing to install, and it can't be removed by accident.
    • One honest limit: the native macOS menu bar stays English for now.
    • Agent mode now runs on the newest Claude tooling, and the model picker gains Fable 5.1 alongside Opus, Sonnet, and Haiku — which now point at the latest version of each model automatically.
    • Claude's task list keeps working with the new tooling: the TODO panel fills in and updates as Claude plans and checks off its work.
    • If you closed a terminal session while a program inside it was still running, the session could come back — often after your computer woke from sleep — as a black, unusable entry in the sidebar that no amount of closing would remove. Closed sessions now stay gone, and whatever was still running inside them is shut down with them.
    • Agent sessions you didn't name are now named after your first message, so the sidebar reads like a list of what you're working on instead of "Session 1", "Session 2". Rename any of them whenever you like.
    • The usage panel now counts input tokens from the whole session, including the turns that happened before you opened the panel.
    • The Context Panel action on the start screen works before you've opened a session.
    • See the full changelog at https://hermes-ide.com/changelog

    Fixed

    • Switching to another session and back now opens the conversation at the latest message instead of at the top.
    • Split Right and Split Down from the terminal's right-click menu now open a fresh session with a working prompt instead of an empty pane.
    • On Windows, Hermes no longer opens off-screen or "vanishes" after being minimized or maximized.
    • On Windows, agent sessions no longer crash on the first message with an EISDIR: lstat 'C:' error.
    • Session cards and pane headers show the model you're actually using instead of just "claude".
    • When Hermes can't create a session's branch workspace, you now get a clear message instead of the session silently not appearing.
    • Keystrokes typed inside full-screen programs (Claude, vim, less, and the like) no longer end up in your shell history and suggestions.
    • Kiro's auto mode now launches with the right permission setting.

  4. v1.3.2

    May 18, 2026

    Improved

    • Recover agent sessions from prompt-too-long; route /compact natively

  5. v1.3.1

    May 16, 2026

    Fixed

    • Fix subagent popover surface contrast and frosted-theme consistency

  6. v1.3.0

    May 16, 2026

    Improved

    • Branch picker single-click + agent masthead ticker polish

  7. v1.2.4

    May 13, 2026

    Fixed

    • Fix model/permission/effort chips disappearing after session switch

  8. v1.2.3

    May 13, 2026

    Fixed

    • Fix branch-selector re-expand panel collapsing immediately
    • Fix four critical agent-mode branch / worktree regressions (1.2.x)

  9. v1.2.2

    May 12, 2026

    Improved

    • Working-state redesign, sub-agent visibility, TODO drift, bug fixes

  10. v1.2.1

    May 11, 2026

    New

    • A focused stability release closing six regressions introduced by the 1.2 design refresh.
    • Typing /remote-control random, /agents create foo, or any other CLI slash command followed by arguments used to be rejected with "isn't available in this environment." These commands now open the embedded terminal with their arguments intact, the way they always should have.
    • Running /mcp, /agents, /remote-control, etc. from a Finder- or Launchpad-launched Hermes was failing with "No viable candidates found in PATH" because the system PATH the app inherits at launch doesn't include the toolchain folders where claude actually lives. The embedded terminal now augments its PATH the same way the main agent does, so the binary is found regardless of how you launched the app.
    • The thumbnail row above the composer now has proper breathing room below the images. Previously the bottom of an attached thumbnail sat flush against the text input.
    • Two distinct bugs were causing the expand button on long code blocks to "move and move back" — leaving you stabbing at it three or four times before a click registered. Both are gone: the button now sits stably under your cursor on press and the first click expands the block.
    • Hitting Stop before Claude streamed its first reply used to leave the "awaiting claude" indicator spinning forever, even though the request was already cancelled and "[Request interrupted by user]" was visible. The indicator now clears the moment the interrupt completes.
    • The Scan button in the Project Context step was hanging off the right edge of narrow dialogs. The footer row now fits regardless of the dialog width — the input shrinks before the buttons do.

  11. v1.2.0

    May 11, 2026

    Improved

    • Design system v2: voice colors, composer refresh, status bar overhaul

  12. v1.1.15

    May 10, 2026

    Improved

    • Theme system v2: 30 themes -> 8, with safe migration for stored ids

  13. v1.1.14

    May 10, 2026

    Improved

    • Right-rail Workbench: per-session Files / Context / Notes panel for agent sessions

  14. v1.1.13

    May 10, 2026

    Fixed

    • Fix composer becoming unresponsive after an IME composition is interrupted

  15. v1.1.12

    May 10, 2026

    Improved

    • Performance audit + Bypass mid-session fix + faster New Session

  16. v1.1.10

    May 10, 2026

    Improved

    • Show streamed thinking text + align first line of code blocks

  17. v1.1.8

    May 10, 2026

    Improved

    • Make the empty-state hero responsive on small displays
    • Make Bypass mode chip apply immediately to in-flight turn

    Fixed

    • Fix update dialog showing no progress feedback during install

  18. v1.1.6

    May 10, 2026

    New

    • Add expand-to-fullscreen view for mermaid diagrams and tables

  19. v1.1.1

    May 9, 2026

    New

    • A small patch release that restores the Windows installer for v1.1, plus a clearer statement of how each platform is supported going forward.
    • Windows installer is back. The v1.1.0 release couldn't produce a Windows build due to a packaging issue. v1.1.1 ships a working _x64-setup.exe (and ARM64 build). macOS and Linux installers were unaffected.
    • Hermes is built and tested primarily on macOS and Linux. Both receive every update.
    • Windows is supported on a best-effort basis. Core terminal features are stable, but newer capabilities — such as Agent mode, introduced in v1.1 — may arrive late on Windows or stay macOS / Linux only. If you rely on Windows day-to-day, pin to a known-good version.
    • The download page on hermes-ide.com now reflects this directly when you select the Windows tab.

  20. v1.1.0

    May 9, 2026

    New

    • The agent timeline got a top-to-bottom redesign that reads like a real conversation, not an instrument-panel logbook.
    • A small avatar chip identifies who's talking — You with your accent color, Hermes with a friendly bot icon. Replaces the cramped № 01 left-gutter numbering.
    • Body text moves to a refined sans-serif at a comfortable reading size; mono is preserved where it matters (code, file paths, tool calls, the cost meter).
    • Your messages appear in a soft accent-tinted card instead of a vertical brass stripe.
    • Whitespace replaces the hairline rules between turns. Conversations breathe.
    • The masthead is calmer — Agent · model · cwd in normal case, no more all-caps tracked mono.
    • Every theme paints this layout differently — hacker keeps its phosphor scanlines, designer adds a subtle paper grain, tron pulses a cyan rail, nightowl gets a glass aurora. Switching themes now changes how a conversation feels, not just what color it is.
    • Type / and Hermes shows every Claude Code slash command in the popover — built-ins, plugins, skills, your custom ~/.claude/commands files. Each entry is clearly labeled:
    • ✦ in-app — runs in the chat (sent as a prompt to Claude)
    • ▣ terminal — opens an embedded terminal that runs claude /<command> interactively
    • Pick /mcp, /agents, /cost, /help, /login, or any other interactive built-in and a small inline terminal pops above the composer running the actual Claude REPL with the command auto-typed for you. Arrow-navigate the TUI, hit Enter, close when done.
    • A new Terminal button next to Builder opens a quick shell prompt right where the chat composer lives — for git status, npm run dev, ls, anything you'd reach for without leaving the agent surface. Same xterm, your default shell, current project as the working directory.
    • Click any MCP server in the right panel and you see:
    • A color-coded status note (Connected / Needs auth / Failed / Unknown) explaining why the dot is the color it is
    • The transport (stdio / sse / http) plus command + args or URL
    • The environment-variable keys the server expects (values are never shown — they may carry secrets)
    • The list of tools that server exposes
    • Restart and Remove actions, with a confirmation step on remove
    • Removing an MCP server now actually disappears it from the panel immediately; cloud-managed servers (claude.ai Gmail, Drive, Calendar, etc.) are detected and labeled "managed elsewhere" so you don't get stuck trying to delete them locally.
    • The "approve / deny / always allow" buttons that appear when Claude asks to run a tool are now real, prominent buttons — not tiny mono-link text. The primary "Approve once" gets a brief attract pulse so your eye lands on the right thing. Standard dialog layout: deny on the left, confirm on the right, Enter submits.
    • Plan mode used to silently drop your responses — clicking Approve on a plan or answering a multi-question prompt looked like nothing happened. Both flows now route through the correct path and Claude sees your answer.
    • Switching the model or permission mode mid-session also stays in sync now; the picker chips reflect Claude's reality even when the change came from inside a conversation (e.g. Claude entered plan mode itself).
    • The Cost & Tokens panel section is gone — its only button didn't work; the running cost meter in the header is the live source of truth
    • The activity-bar Context button now actually toggles the agent context panel (Cmd+E to hide it for more horizontal room)
    • The dirty-close session dialog stops surfacing .aider.chat.history.md and other auto-generated noise files
    • Composer chips truncate cleanly on narrow widths instead of overlapping each other
    • The bot/person icons in the speaker chip take the theme accent — green-phosphor on hacker, cyan on tron, terracotta on designer
    • If the new timeline isn't for you, Settings → Appearance → Use classic compact timeline restores the denser, mono-bodied logbook style this release replaced — with the brass left bar on user messages, hairline rules between turns, and the prior typography. Toggle it back any time. Themes still apply on top.
    • New for 1.0.0: Claude sessions now open in Agent mode by default — a real chat experience with rich tool calls, thinking blocks, and proper images instead of a terminal pretending to be a chat.
    • Talk to Claude in a real chat interface — no more typing into a terminal disguise
    • See Claude's thinking, file edits, and tool results as first-class cards instead of buried in scrollback
    • File edits arrive as proper diff cards you can scan at a glance
    • A clear summary appears at the end of each turn with token usage and timing
    • Paste or drop images that Claude actually sees — not file paths typed for you
    • Slash commands and the model picker pull from Claude itself, so they stay in sync as Claude updates
    • Sessions persist across restarts — reopen and continue where you left off
    • Anything that isn't Claude — Aider, Codex, Gemini, Copilot, Kiro, plain shells — keeps working in classic Terminal mode, unchanged. You can also explicitly open a Claude session as a terminal (for the claude CLI itself, scripting, or muscle memory) by choosing "Open as terminal instead" in the new-session creator.
    • You can convert a session between Agent and Terminal mode with right-click → Convert. Conversation history doesn't carry over, so it's a deliberate switch.
    • Agent mode requires claude (the Claude Code CLI) on your PATH. Hermes detects it on startup and offers to install if missing. Use whatever auth you've already configured for Claude — Pro, Max, or API key all work; Hermes never asks for tokens.
    • Existing saved sessions from 0.6.x continue to open in Terminal mode, exactly as before — nothing changes for restored workspaces
    • New Claude sessions default to Agent mode; uncheck "Open as terminal instead" in the new-session modal to get the old TUI experience
    • Switching modes on an existing session starts a fresh conversation; the previous transcript stays in the original session

    Improved

    • Slash command suggestions for Claude come straight from Claude, so new commands appear automatically as you upgrade
    • The model picker shows Claude's actual current model and effort options instead of a hardcoded list
    • Mentions and image attachments are now part of the chat composer, with previews before you send

  21. v0.6.16

    April 22, 2026

    Improved

    • Per-agent launch-command prefix

    Fixed

    • Fix AI CLI 'Not detected' for tool-manager installs

  22. v0.3.37

    March 8, 2026

    Improved

    • Fresh new app icon with a modern winged H design
    • Updated website favicon to match the new icon

  23. v0.3.36

    March 8, 2026

    New

    • Drag-and-drop images onto AI terminals — the image is copied to the system clipboard for AI vision, and the file path is inserted automatically

    Fixed

    • Closed sessions no longer reappear after restarting the app
    • App no longer crashes on launch due to analytics plugin initialization

  24. v0.3.35

    March 8, 2026

    New

    • Session restore on relaunch — all open sessions are automatically saved and restored when the app is closed and reopened, including scrollback content, layout, colors, descriptions, projects, and groups
    • "Restore Sessions on Launch" setting — choose to always restore your previous workspace or start fresh every time
    • SEO Audit & Optimization prompt template — run a full SEO audit covering meta tags, Open Graph, structured data, sitemap, and keyword strategy
    • Landing Page SEO Copy prompt template — generate ready-to-use SEO copy with titles, descriptions, headlines, CTAs, and alt text

  25. v0.3.34

    March 8, 2026

    New

    • Auto-approve option in the New Session wizard — launch Claude, Gemini, Aider, or Codex in fully autonomous mode with a single checkbox
    • First-launch onboarding wizard — choose a theme, review what data is collected, and accept the privacy policy before getting started
    • "Needs input" session status — sessions now show when an AI agent is waiting for your confirmation or approval, distinct from idle and working states
    • Cancel button for in-progress updates — stop a download anytime and retry later, with download size info and slow-connection warning
    • Empty projects — create a project before adding any sessions, so you can organize your workspace upfront
    • Drag-and-drop sessions between projects — move sessions into a project, out of a project, or between projects by dragging in the sidebar

    Improved

    • Smarter session activity detection — uses output silence timing instead of brittle text matching, so idle, working, and needs-input states are accurate even with interactive permission dialogs
    • Color band on sessions and projects is now easier to spot and click when no color is assigned — shows a dashed outline that highlights on hover
    • Creating a session from a project's "+" button now pre-selects that project in the New Session wizard
    • New Project input moved inline into the sidebar footer for a cleaner, more predictable position

    Fixed

    • Closing a session no longer freezes the app when the terminal process is unresponsive
    • Zombie sessions that stop responding can now be force-closed from the sidebar
    • Force-close fallback for zombie sessions that the backend can't clean up

  26. v0.3.33

    March 7, 2026

    New

    • Optional session colors — sessions start with no color by default, and you can add one anytime from a 12-color palette or a custom hex value
    • Terminal background tint — when a session has a color, the terminal shows a subtle tinted overlay for quick visual identification
    • Color picker with preset palette, custom hex input, and a "None" option to remove color
    • Project-wide color changes with confirmation — changing a project's color prompts before applying to all sessions
    • Choose a color when creating a new session, or leave it blank
    • Session description field — click to add or edit a short description for any session
    • Inline renaming for sessions and projects — double-click any name to rename it directly in the sidebar
    • Visual project management in the sidebar — create, rename, and organize projects without right-clicking
    • Assign sessions to a project directly from the session creation wizard
    • Download icon next to the version number for quick access to updates

    Improved

    • "Groups" renamed to "Projects" and workspace "Projects" renamed to "Folders" for clearer terminology
    • Moving a session into a project automatically adopts the project's color
    • Color bands replace session numbers and dots for a cleaner sidebar
    • Terminal panes show a colored left border matching the session's color
    • Hidden files (like .env) are now visible by default in the file explorer
    • Session label, description, and color changes are now persisted to disk

    Removed

    • Telemetry and analytics removed entirely from the app

  27. v0.3.32

    March 7, 2026

    Improved

    • Click the version number in the status bar to manually check for updates
    • App icon now displays at the correct size in the macOS Dock

  28. v0.3.31

    March 7, 2026

    New

    • "What's New" dialog after app updates — shows what changed in the new version with an option to permanently dismiss
    • Redesigned left navigation — Git, Files, and Search are now nested inside each session, making the hierarchy clear
    • Session sub-view buttons appear directly under the active session and stay accessible while browsing
    • Drag-and-drop files directly into terminal sessions to insert their paths
    • Working directory path indicator in the Git panel
    • "Isolated copy" badge for sessions using git worktrees
    • Hover tooltips added across all buttons and interactive elements

  29. v0.3.30

    March 5, 2026

    New

    • Drag-and-drop files directly into terminal sessions to insert their paths

    Fixed

    • Dual-selection bug in the New Session wizard AI Engine step

  30. v0.3.29

    March 3, 2026

    New

    • Redesigned command toolbar — compact single-row bar with a Compose button, quick-action pills, and a searchable Commands dropdown
    • Commands dropdown with search/filter, category grouping, and full keyboard navigation (arrow keys, Enter, Escape)
    • Quick actions show your recently used commands plus smart defaults for each AI provider
    • Settings shortcut at the bottom of the left activity bar
    • Bug report button in the status bar — opens a pre-filled issue with your app version, OS, and OS version
    • Bug reports now use GitHub Issues — no sign-in wall, drag-and-drop screenshots, and structured fields

    Improved

    • Prompt Composer is now accessible directly from the command toolbar Compose button
    • Commands dropdown auto-hides when a provider has 5 or fewer commands (e.g., Copilot shows both inline)
    • OS version auto-detected when filing bug reports

  31. v0.3.28

    March 3, 2026

    New

    • Full keyboard navigation for the New Session wizard — arrow keys to navigate, Space to toggle, Enter to advance
    • Keyboard shortcut hints shown at the bottom of each wizard step
    • Shell picker in Settings now dynamically detects installed shells on your machine
    • Cross-platform shell detection: supports zsh, bash, fish, nu, sh on Unix and PowerShell, cmd.exe, Git Bash on Windows

    Fixed

    • Changing the default shell in Settings now takes effect when opening a new terminal
    • Dead key composition (ã, õ, é, etc.) no longer triggers unexpected behavior in bash

  32. v0.3.27

    March 3, 2026

    Fixed

    • Dead key composition fully reworked for reliable international character input
    • Typing contractions like "don't" no longer produces double apostrophes or drops characters
    • Combining dead keys (ã, é, ô, ü, etc.) now display correctly in all terminal sessions

  33. v0.3.26

    March 2, 2026

    Fixed

    • Text duplication after dead key composition in terminal

  34. v0.3.25

    March 2, 2026

    Fixed

    • Fast typing no longer causes missing characters or a stuck terminal
    • Short character bursts during fast typing are no longer incorrectly suppressed

  35. v0.3.24

    March 1, 2026

    New

    • Pinnable templates — pin frequently-used templates so they appear in a dedicated "Pinned" section at the top of the template picker
    • Pin state persists across sessions and is included in settings export/import

  36. v0.3.23

    March 1, 2026

    New

    • Legal & Compliance template category with 8 templates: Privacy Policy, Terms of Service, Legal Document Review, Contract Clause Analysis, GDPR/Data Protection Audit, Software License Advisor, EULA Generator, and Cookie & Consent Policy (105 total templates across 18 categories)
    • Legal Advisor role for privacy regulations, contract drafting, software licensing, and compliance
    • Bug report button on the website

    Fixed

    • Text duplication while typing in terminal when AI agent is outputting

  37. v0.3.22

    March 1, 2026

    Fixed

    • Text duplication when typing while an AI agent is actively outputting
    • False-positive cost detection on long code output lines
    • Sessions no longer flash a "destroyed" state when closing — they disappear cleanly
    • All panels now collapse when the last session is closed
    • Linux x86_64 AppImage build fixed

  38. v0.3.20

    March 1, 2026

    New

    • Full support for all 5 AI providers: Claude, Gemini, Aider, Codex, and Copilot — each with auto-detection, token tracking, cost monitoring, and native slash commands
    • 45 new prompt templates across 6 new categories (97 total across 17 categories): Product, Planning, Design, DevOps, Data, AI/ML, Incident Response, Business, and Monetization
    • 80s theme — authentic amber CRT retro aesthetic with beveled 3D controls, phosphor glow, and scanline overlay
    • Duel theme — red vs green lightsaber energy effects with ambient glow
    • Rainbow theme — spectrum energy lines with ambient color glow
    • Theme picker in the status bar for quick switching between 11 themes
    • Split download and install for auto-updater — download in background, install when ready

    Fixed

    • Agent detection now works correctly when manually typing provider commands
    • Auto-updater: dismissed dialog re-shows when a newer version is found
    • Auto-updater: download and install failures now show proper error feedback
    • Auto-updater: prevented double-downloads and concurrent race conditions
    • 80s theme sharp corners preserved correctly when changing UI scale
    • Model name detection no longer false-matches on similar substrings
    • Theme picker no longer hidden behind scanline overlay

    Improved

    • Tron set as default theme with animated energy line effects
    • Warm purple undertone for Rainbow theme backgrounds

  39. v0.3.19

    February 28, 2026

    New

    • Lightweight privacy-friendly analytics (opt-out in Settings > Privacy)
    • Privacy tab in Settings with telemetry toggle
    • Settings sections searchable in command palette

    Fixed

    • Checkbox spacing in settings label rows
    • Analytics correctly reports to production in release builds

  40. v0.3.17

    February 28, 2026

    Fixed

    • Terminal copy now works correctly
    • 17 bugs fixed across the app
    • Platform detection no longer misidentifies macOS as Windows
    • Context menus properly cleaned up on dismiss
    • Prompt Composer stays open on send failure so you can retry

  41. v0.3.14

    February 28, 2026

    New

    • ARM64 builds for Linux and Windows
    • Dynamic download links on the website — always points to the latest release

  42. v0.3.9

    February 20, 2026

    New

    • Native macOS menu bar with full keyboard shortcut coverage
    • App-wide context menus (right-click) for sessions, files, git, and more

    Fixed

    • 34 bugs across all subsystems
    • Prompt Composer no longer sends a spurious Enter before the composed prompt

  43. v0.3.8

    February 15, 2026

    Fixed

    • 45 bugs fixed across all subsystems