Desktop: a janela inteira era redesenhada a cada arquivo salvo. Cada arquivo que um agente escrevia redesenhava os painéis e as linhas da sidebar de todos os projetos abertos, e com vários agentes trabalhando o app podia parar de responder. Agora só redesenha quando algo visível mudou.
Desktop: o painel de alterações de uma worktree atualiza enquanto o agente edita, não só depois de um commit.
Desktop: worktrees. Um agente pode trabalhar num checkout separado do projeto, na própria branch, sem mexer na pasta original. Dois agentes podem dividir uma worktree ou cada um ter a sua. Vem desligado: liga em Configurações → Agentes. Com um projeto Git aberto, um aviso na sidebar oferece ligar.
Desktop: aba Worktrees no painel lateral. Crie uma numa branch nova a partir de qualquer ponto, ou numa branch que já existe. Se o projeto tem uma preparação configurada (arquivos para copiar, como o .env, e comandos para rodar, como npm install), ela roda na criação ou depois. A preparação fica nas configurações do projeto, em Worktrees.
Desktop: mova um agente para uma worktree, ou duplique ele lá, pelo menu do agente. Se estiver rodando, ele é parado e depois movido. Cada lugar guarda a própria conversa, então o agente que volta para a pasta original retoma de onde parou ali.
Desktop: a sidebar agrupa os agentes por onde eles trabalham assim que uma worktree entra em uso: um bloco para a branch original e um por worktree. Botão direito no cabeçalho do bloco adiciona um agente ou abre um terminal naquela pasta. Arquivos, alterações e Source Control mostram a pasta do agente que você está olhando.
Desktop: faça o merge de uma worktree pelo painel. Ele lista os commits que vão entrar e o que está atrapalhando, como mudanças não commitadas ou um agente rodando. Conflito manda você para o Source Control resolver. Quando a branch já está na original, o bloco avisa e oferece limpar: remover a pasta, apagar a branch e mover ou remover os agentes que moravam ali.
Desktop: os agentes fazem o mesmo por MCP. Criar e listar worktrees, mover um colega, fazer merge e remover, ou já criar um agente novo dentro de uma worktree. Rotina também pode rodar numa worktree.
Corrigido
4
Desktop: um pull recusado por mudanças não commitadas diz isso, em vez de dizer que a branch divergiu. Faça commit ou descarte as mudanças, depois puxe de novo.
Desktop: pasta de trabalho relativa no hive.yml é resolvida a partir da pasta do projeto, não de onde o app foi aberto. Um agente iniciado por outro agente sem pasta de trabalho também roda na pasta do projeto.
Desktop: a lista de branches não mostra mais uma linha "+ branch" perdida.
Desktop: um agente reiniciado fica com a conversa nova. Um sinal atrasado da execução que acabou de terminar não sobrescreve mais.
Desktop: recorte a foto antes de ela virar avatar. O ícone do workspace, a foto da conta e o avatar do agente abrem uma janela de recorte: arraste para posicionar, role ou deslize para aproximar, e veja o resultado no tamanho em que ele vai aparecer.
Desktop: reordene projetos arrastando o cabeçalho. Segure e mova, como os círculos de workspace. O ícone de arrastar saiu.
Desktop: um agente pode trabalhar com os agentes de outra hive aberta. Listar, iniciar, parar e mandar mensagem aceitam um agente de outro projeto, desde que o id o nomeie. Um nome solto é recusado, porque com várias hives abertas "Claude" é vários agentes.
Desktop: um agente pode retomar a conversa de um colega em vez de só começar uma nova. Recusa quando não há sessão salva, em vez de recomeçar sem avisar.
Desktop: o chat mostra o rascunho de feedback que o Claude Code escreve sobre si mesmo, com os detalhes em Revisar e os mesmos Enviar e Descartar que o terminal oferece. Nada é enviado sem você.
Mudou
2
Desktop: "Discutir isso" é uma ação abaixo do card de pergunta, não uma escolha numerada. Os números são só as opções do agente.
Desktop: ler uma conversa longa. O card de aprovação devolve a tela enquanto você sobe para ler o plano que ele pergunta. Trocar de agente mantém sua posição. O histórico se completa sozinho em vez de esperar um clique.
Corrigido
12
Desktop: Codex no Windows. Mensagens de versões do Codex anteriores à 0.147 aparecem no chat; antes apareciam só como "Trabalhou por Ns". Enter no composer envia a mensagem em vez de quebrar linha.
Desktop: mensagem enviada com o agente ocupado não some mais, nem deixa o chat parecendo travado com "Trabalhando…" e um balão que nunca assenta.
Desktop: um prompt que o agente não recebeu volta para o composer. Quando o Codex recusa uma mensagem no limite de uso, ou sai logo depois de você enviar (uma atualização, uma queda), o texto volta para a caixa de mensagem em vez de sumir. O aviso de limite fica.
Desktop: o card de pergunta mostra a pergunta inteira, não só a última linha, e não a repete acima do card nem mostra a faixa de abas como texto.
Desktop: um menu citado dentro de uma resposta não é uma pergunta. Quando o agente colou o "Yes, continue / No, quit" de outro agente e continuou trabalhando, o chat desenhava um card e travava o composer.
Desktop: o consentimento de bypass de permissões na abertura do Claude aparece como pergunta em vez de o agente parecer travado.
Desktop: um card de pergunta já respondido não volta mais para o fim do chat, e uma mensagem com imagem não aparece duas vezes.
Desktop: um agente criado por outro agente inicia com segurança. Uma mensagem entregue enquanto o terminal dele ainda abria caía no lugar errado; numa pasta que o Codex nunca tinha visto, caía no prompt "Do you trust this directory?" e fechava o agente. A entrega agora espera o terminal ficar pronto e nunca digita nesse prompt. Responder a ele continua sendo seu.
Desktop: o chip de uso do Codex no Windows sumia quando o comando do agente era um caminho entre aspas.
Desktop: uma janela semanal esgotada é a que o chip de uso mostra, mesmo com a janela de 5 horas sobrando.
Desktop: layouts divididos sobrevivem a um reinício.
Desktop: um agente fixado mantém a marca do provedor e o arrasto.
Desktop: times na sidebar. Junte agentes em grupos com nome, arraste para reordenar, renomeie o time ali mesmo e recolha os que você não está olhando. Renomear um time não desfaz o que você tinha recolhido.
Desktop: o cabeçalho nomeia cada painel do split. A barra de cima mostra uma etiqueta por agente na tela, e não só a do painel em foco. Clique na etiqueta para ir até o painel.
Desktop: um mascote do Claude no composer. No chat de um agente Claude, um Claude pixelado dorme na borda de cima da caixa de mensagem e reage ao agente: pula para fora quando você envia, toma um café enquanto ele espera por você, e se mexe mais conforme o turno se estende. Dá para desligar em Configurações › Agentes.
O changelog passa a sair em português e espanhol, em hiveterm.com/pt/changelog e hiveterm.com/es/changelog.
Corrigido
8
Desktop: o histórico do chat rola enquanto o agente ainda escreve. No chat do Codex ele travava no meio da resposta.
Desktop: várias imagens numa mensagem ficam em uma linha, em vez de empilhar torto.
Desktop: dá para fechar o último painel do split com a barra lateral recolhida. A prévia que abre no hover ficava em cima do botão de fechar.
Desktop: a mensagem que você acabou de enviar já nasce no lugar, em vez de aparecer na beirada de baixo e pular para o lugar um instante depois.
Desktop: escolher um template de instrução pronto não apaga mais o agente que você tinha escolhido.
Desktop: um modelo esgotado não fala mais pela conta inteira. Com um modelo no limite, o chip de uso marcava 100% enquanto a janela de cinco horas ainda tinha quase tudo livre. Agora o chip mostra a janela da conta, e o modelo esgotado continua com a linha dele quando você abre os detalhes.
Desktop: as barras de baixo ficam todas na mesma altura, e a foto da conta alinha com o "Add project".
Desktop: o chat do Codex, conferido contra uma sessão de verdade. Sub-agentes aparecem como card e param quando param. Duas mensagens iguais na fila continuam duas. O /new segue a sessão em que você está, e dois agentes na mesma pasta não pegam mais a do outro. Uma linha de comando enorme continua aparecendo, a compactação desenha um divisor só, o texto do plano não some, e formulário com anotação é lido direito.
Segurança
1
Desktop: um número de processo inválido não alcança mais nada além do agente. Parar um agente com um número fora de faixa podia sinalizar todo processo do seu usuário. Esses números são recusados agora.
Desktop: context window ring. In the chat composer and the status bar, a ring fills as the conversation grows, for Claude, Codex and Grok, reading the same number each CLI shows on its own screen. Click it for the numbers behind it: used and free, what the last turn cost (how much came from cache), what lives in the window, and what happens near the limit.
Desktop: an amber pin on the agent's row when it finished a turn or is waiting on you while you were looking at something else. It clears the moment you open the agent.
Mudou
8
Desktop: messages between agents arrive. A message or a reply sent to a live agent (Claude, Codex, Grok, Antigravity) now reaches its session instead of sitting in an inbox it never reads; the sender is told how it was delivered.
Desktop: an agent created by another agent no longer takes your focus. It appears in the list with the pin, opens in the view its creator was using (chat or terminal), and starts in the background.
Desktop: a question from another agent is titled after the question, not after "Hive ask from…".
Desktop: creating an agent from an agent names the agent CLI (claude, codex, grok, agy…), never a shell command — processes have their own tool. Claude Code's auto mode used to block the call as if it were a shell command; HiveTerm now pre-approves its own tools there, except running commands, typing into another agent and deleting an agent, which still ask.
Desktop: the sidebar preview is the agent's last reply, not the end-of-turn summary line.
Desktop: ⌘1–9 follow the order you see in the sidebar, which sorts agents by who replied last.
Desktop: the reconnecting notice sits in the header, at the header's height, instead of a strip over the title bar.
Desktop: the plan toggle sits after the effort chip in the Codex composer, and the usage popover names the provider you need to sign in to instead of saying Claude Code for everyone.
Corrigido
7
Desktop: a Claude started inside an agent could take over the agent's chat. A probe, a test suite, or a claude -p an agent ran from a tool made the chat show that stray conversation and pointed Resume at it. Only the agent's own process speaks for it now.
Desktop: accents corrupted in the chat — a "—" or a "ç" cut at a read boundary showed as "�".
Desktop: "Load earlier" came back empty in Codex chats, and a message with a pasted screenshot could vanish from earlier history.
Desktop: two Codex agents in the same project could show each other's conversation, and a new agent could pick up the conversation of one removed earlier with the same name.
Desktop: a question sent to a Codex agent by another agent could come back without the answer.
Desktop: Grok agents were renamed to "- Waiting for response…"; only the conversation's summary names the row now.
Desktop: the effort chip was empty at the start of a Claude session; it reads the per-model effort Claude Code saves.
Desktop: Codex in the chat view. Everything the chat does for Claude now works for Codex: your messages and the agent's replies, the commands it runs, images and files, and the answer you owe it — an approval, the questions it asks while planning (each one a card, with a note if you want), and the "Implement this plan?" decision at the end of a plan, which is shown as a plan card. Model and effort switch from the composer using Codex's own picker, with the models your account actually has; two chips show Plan mode and the permission level. Queued messages, Stop, and picking a past conversation with /resume work like they do for Claude.
Desktop: plan usage for Codex, Grok and Antigravity. The footer shows how much of your plan is left, with the weekly window for Claude, for accounts signed in through ChatGPT, Grok and Google.
Desktop: agents can run the hive. New tools let an agent start, stop, interrupt, restart or remove another agent, set its instructions, create a temporary helper or a permanent agent, and run a shell command — the same actions as the sidebar buttons, with the same names you see on screen.
Desktop: images and files in one message. Attach several, write around them, and the chat keeps each reference where you put it, with numbered thumbnails and file cards that match.
Desktop: no dead ends around a missing CLI. Picking an agent whose CLI is not installed shows the install command to copy and a link to its docs, in the add-agent dialog, in onboarding and in the terminal.
Desktop: provider marks in colour in the add-agent dialog and onboarding; agents that are not installed sit on a second row, greyed out, and cannot be picked by mistake.
Mudou
5
Desktop: Codex instructions load with the agent. They are part of the agent's setup now, survive long sessions, and no longer appear as a first message in the chat. An edit takes effect on the next start.
Desktop: clicking the agent's name in the header focuses the agent when a file is open; a second click opens its settings as before.
Desktop: the "Welcome to Pro" card only appears when you actually upgrade, not when a Pro account signs in on a fresh install.
Desktop: clearer wording in the first-run tips about restarting agents.
Mobile: provider marks match the desktop — the same vector marks and brand colours on agent rows, top bars and the spawn sheet; a stopped agent fades to grey.
Corrigido
6
Desktop: chat messages you queued in the terminal sometimes vanished from the Claude chat; they stay listed until the agent takes them.
Desktop: a message with several images and a line break after an attachment did not reach the agent; each attachment is now sent on its own.
Desktop: the chat stayed on "Working…" after you interrupted Codex, and a whole conversation could disappear after the agent used a Hive tool.
Desktop: switching to the dark theme applies immediately.
Desktop: the effort chip cycles through the levels each Codex model supports, including Ultra where available, and follows the level Codex sets per mode.
Desktop: fewer macOS password prompts when the app reads provider sign-ins from the keychain.
Desktop: a workspace column. Your workspaces now live in a slim column left of the sidebar, one circle each: click to switch, right-click to rename or give it an image, and the + creates a new one right there — name and photo in a small card beside the button. Your account moved to the bottom of that column, with "Change photo…" in its menu so it can wear your face too.
Desktop: a command palette worth opening. Bigger, with sections — your agents first, then agents from other workspaces, then actions and files — icons on every action, key hints at the bottom, and a placeholder that says what it does. Agents show their avatar and project; Stop is a button on the row instead of a look-alike entry; picking an agent from another workspace takes you there.
Desktop: an editor that behaves like the one you know. Find and replace (⌘F) with whole-word, regex and multi-line search, replace one or all, bracket matching and auto-closing, code folding, indentation guides, multiple cursors and VS Code shortcuts. A folder button reveals the open file on disk.
Desktop: the chat shows what the agent is doing with your settings. Switching model or effort from inside the chat updates the composer chip; each reply ends with when it finished ("done 5:59 PM"); Claude's own recap appears as a recap card, queued messages and compactions are shown, and a tool the agent was refused is marked as such. Fable 5.1 is in the model list.
Desktop: the last message under Antigravity agents too, like Claude, Codex, Grok and OpenCode already had.
Desktop: Clone a repository has a shortcut (⌘⇧O), editable in Settings, and ⌘O sits on "Open local folder…" where it belongs.
Desktop: real tooltips with the shortcut inside on the workspace column, the sidebar toggles and the right column's toolbar.
Mudou
10
Desktop: switching workspaces brings you back where you were. Leaving a workspace remembers the agent you were talking to; returning puts it back on screen — in both directions, and after a relaunch.
Desktop: a leaner sidebar. The filter field and the workspaces section are gone (the palette and the column took over), the collapse and expand buttons are arrows that stay in their corner, and an empty workspace shows one clear "Add project" card instead of a lonely line.
Desktop: "Add agent" fits the window and its Instructions field is always open.
Desktop: the first-run tour starts at workspaces, then the sidebar, the agent's column and routines; a terminal-only run skips the tips about agents. Tooltips stay quiet while a tip is up.
Desktop: "Customize an agent" also checks itself off when the agent picks its own avatar through the Hive tools.
Desktop: composer buttons are proper circles, the Clear button only shows on terminals and commands, and there is no pointer cursor or text cursor over sidebar rows anymore.
Mobile: the app opens on your projects. The home screen is gone, and with it a whole level of tapping — opening the app and reaching an agent is two taps now. Each project says in one line whether it needs you: what's running, what's waiting for you, how many changes are uncommitted. Your desktops moved to the button in the top-left corner.
Mobile: a project opens as a list of its agents. Every agent shows its picture, the last thing it said and when it said it, so a glance tells you who has been working and who is waiting on your answer. The agents that need you sit at the top.
Mobile: an agent's routines travel with the agent. Following the desktop, a routine now lives inside the agent that runs it, next to its terminal — and creating one from there stops asking which project and which agent, because you are already in them. Routines that belong to no agent stay with their project.
Mobile: the app looks like the desktop again. The new colors, the same icons, and the same controls in the same place on every screen — including the back button, which no longer changes shape or position between screens.
Corrigido
7
Desktop: agents you added during a session now come back after closing the app — for every CLI, not only Claude. Codex and Antigravity sessions in particular no longer start over; a Codex agent that had not sent its first message yet is remembered as soon as it does.
Desktop: two Codex agents in one project no longer steal each other's preview.
Desktop: the first message to a brand-new agent no longer stays frozen under its reply.
Desktop: the chat no longer repeats the agent's reply as a summary, and Claude's recap no longer shows twice.
Desktop: text selection in the terminal works again, and copying from the chat keeps working.
Desktop: opening the chat with many agents is much lighter, which is where the freezes on Windows came from.
Desktop: the hover on a collapsed project's chevron is centered, and a pinned agent card shows its Start, Stop and Restart actions on hover.
Desktop: the first-run tour follows the new layout. The tips now walk through the three things that matter — your project's agents in the sidebar, the agent's column on the right, and routines — each pointing at the real control, with clearer wording. The "First steps" checklist swaps "Split the view" for "Customize an agent": give one your own instructions or an avatar and it checks itself off.
Corrigido
2
Desktop, Windows: no more stray black console windows. Every Claude Code session could pop a bash window that stayed open behind the app — one per session, piling up over a day. They came from how HiveTerm registered its Claude Code hooks; it now runs them silently. Existing installs are fixed automatically on the next launch, and any windows still open can simply be closed.
Desktop: the chat's effort indicator no longer jumps to "ultracode" on its own. When Claude Code printed a tip mentioning /ultracode, the chat view read it as a change of effort level. Only the real level counts now.
Desktop: pin your favorite agents. Right-click an agent and choose Pin: it becomes a card at the top of the sidebar, avatar front and center with its latest reply underneath. Pin two or more and they tile side by side. Everything you can do on a normal row — rename, settings, split, remove — works on the card too.
Desktop: the agents you talk to rise to the top. Below the pinned cards, agents reorder by their most recent reply, like a chat app — the one you just worked with is always first.
Desktop: the last message under every agent's name. The sidebar shows a one-line preview of each agent's latest reply — now for Codex, Grok and OpenCode too, including sessions you resumed after a restart.
Desktop: custom agent avatars. Give an agent a face: open its settings and click the avatar to pick an image — or ask the agent to choose one for itself. Agents can also rename themselves the proper way now, instead of you editing config files.
Desktop: run-once routines. The schedule picker gained "Once": pick a day and time and the task runs a single time — "review this tomorrow at 15:00". It keeps its run in the history and pauses itself when done. Agents can schedule one-shots for you too.
Desktop: one side column for everything. Routines, Source Control, Files and Search now share the agent's column on the right, switched by the icon strip at the top — the content swaps in place, nothing resizes or jumps.
Desktop: your account in the sidebar footer. Your email sits at the bottom with a menu for Settings, Theme, About and Sign out — and Add project right above it.
Mobile: paste a screenshot into an agent. Paste now takes images, not just text. Copy a screenshot anywhere on your phone and tap Paste: the picture goes to your desktop and the agent gets it, the same way dragging the file onto the desktop app works. The button also stops looking unavailable when the only thing on your clipboard is an image.
Mudou
3
Desktop: a new look across the whole app. Every icon was replaced with a crisper, more distinctive set, corner rounding is consistent everywhere, and the light and dark themes were rebuilt around clean neutral surfaces — color now belongs to your agents and your selections, not to every panel.
Desktop: the agent owns the screen. The focused agent's avatar and name sit at the top of the workspace; open documents appear as tabs beside them; git — branch, sync and the GitHub link — moved down to the status bar. The old tab bar is gone.
Desktop: everything about an agent updates live. Rename it, change its instructions or its avatar — from its settings, from the config file, or by asking the agent itself — and the whole app reflects it instantly, in every project.
Corrigido
3
Desktop: Windows updates stop stumbling through versions. Updating after skipping a few releases could march through one installer error after another. From this release on, updates land directly on the newest version. If updating has been failing for you, uninstall HiveTerm once and grab a fresh installer from the website — after that it stays smooth.
Desktop: resumed agents act like they never left. An agent resumed after a restart now shows its last-message preview, and a routine it runs finishes properly instead of showing "running" forever.
Desktop: footer menus close on a second click. Clicking the account or Add project button while its menu was open used to reopen it instead of closing.
Mobile: paste into an agent's terminal. A Paste button sits in the row of keys below the terminal, next to attach and dictation — tap it to drop whatever is on your clipboard into the agent's prompt. Several lines arrive as a single paste, so an agent takes a pasted error or log as one message instead of running the first line on its own. The button dims when there is nothing to paste, and the key row was reshaped to fit it on smaller phones: the arrow keys are now one up/down control, and the keyboard-hide button moved out of the way of your thumb.
Corrigido
5
Desktop: Restart works on Windows. The Restart button in the sidebar used to freeze the agent instead of restarting it — the only way through was Stop and then Start. Restart now brings the agent back the way it does on macOS and Linux.
Desktop: split terminals no longer garble each other. With two or three agents side by side, the panes you were not looking at could turn into scrambled characters while another one was printing output. Every pane now repaints correctly.
Desktop: sub-agents start clean from the sidebar too. 0.43.5 separated Start from Resume for sub-agents, but the sidebar's own Start and Restart buttons kept continuing the previous conversation. All of them now agree: Start is a fresh conversation, Resume continues the recorded one.
Desktop: a helper agent launched without a task no longer loops. Starting a sub-agent and just saying hello could leave it endlessly asking for work, burning through tokens. It now waits for its task, answers it, and stops.
Desktop: agents start in projects whose folder name has an accent. A project path containing a character like "ã" or "ç" could stop an agent from starting at all.
Desktop: Claude agents with instructions crashed on Resume and at launch. Since 0.43.5, a Claude agent that has instructions set could not continue its conversation — the Resume button and the automatic restore on launch both ended in "Crashed" with an error about session ids. Fresh starts and other agents (Grok, Codex, …) were not affected. Resume works again.
Desktop: Stop ends the whole agent on Windows. Stopping an agent used to leave it running in the background for a while, and the next Start could pick up the old conversation. Stop now takes the agent and everything it launched down at once, and Resume points at the right session.
Desktop: Start means a fresh session, even right after launch. If you pressed Start while HiveTerm was still bringing an agent back from your last session, the resumed conversation could win over your click. Your latest action now always wins — Start, Restart, Resume or Stop.
Desktop: sub-agents start clean. Start on a sub-agent used to continue its previous conversation exactly like Resume; the two buttons now do different things.
Desktop: Chat attaches to a new agent's conversation right away. A freshly started agent no longer needs to report in before Chat knows which conversation to show.
Desktop: consulting a stopped agent works on Windows. An agent asking a stopped colleague a question no longer lands in an empty command prompt.
Desktop: Chat's "Your choices" card stops floating. After answering an agent's questions, the summary card could stay pinned at the bottom for the rest of the session, under every later message. It now settles next to the question once the agent has your answers, and a second round of questions no longer mixes in the previous answers.
Mobile: routines on the phone. Each project lists its routines with schedule, next run and last outcome. Create and edit routines from the phone — pick a project agent, dictate the task, set the schedule — run one on the spot, pause it, or delete it.
Mobile: talk to a routine. Open a routine to read what it did, watch a run as it happens and reply to keep the conversation going, with attachments, dictation and the agent's effort and model at hand. Routines that run inside a project agent open that agent's terminal, as on the desktop.
Desktop: routines stay in sync with your paired phones. Every routine, run and reply reaches the phone the moment it changes.
Corrigido
1
Mobile: tapping a routine notification opens that routine's conversation instead of landing on the home screen.
Desktop: your phone stays quiet while you're at the computer. If you've touched your keyboard or mouse in the last couple of minutes, agent alerts skip the phone — the desktop notification already reached you. Step away and pushes resume. Toggle it in Settings → Notifications.
Corrigido
3
Desktop: clicking a notification opens that agent's session. On macOS a click used to only bring the window forward; now it lands on the bee — or the routine's chat — that pinged you, switching workspace and project if needed.
Mobile: notification taps land reliably. Tapping now opens the agent's terminal even when the tap launches the app or the desktop takes a few seconds to connect — the app finishes connecting and then takes you there.
Desktop: the branch switcher hides when a project has no branches yet. A brand-new repo without commits no longer shows a branch button with an empty menu.
Desktop: first launch walks you through setup. Welcome, pick an agent that's already on your machine, then open a folder, clone a repo, or try a tiny demo page with a bug for the agent to fix. You can skip anytime. After you're in, short tips and a First steps list in the sidebar check off as you go.
Desktop: Chat shows sub-agents while they run. When an agent fans out workers, they sit above the composer with a status mark — you don't have to flip to Terminal to see they're still going.
Corrigido
2
Desktop: canceling a Chat message puts the text back in the box. Like Terminal — a quick cancel doesn't leave a frozen bubble you can't edit.
Desktop: resuming a Chat session lands on the latest messages. It no longer sticks on the empty "What are we building today?" screen.
Desktop: an agent can create and manage routines. Through Hive it can add, list, pause, delete, or run a routine — the same actions as the Routines sidebar. Names match what you see in the sidebar.
Corrigido
2
Desktop: a pasted image in Chat shows its thumbnail right away. You no longer have to switch agents and come back to see the preview.
Desktop: Hive tool calls in Chat no longer overlap as a jumble of names. A call shows as a short chip like hiveterm:update_routine.
Desktop: routines run in the agent's own conversation. Run now lands on that bee — if it's stopped, Hive starts it; if it's busy, Hive interrupts and injects. No more throwaway overlay chat.
Desktop: an agent can add another agent to the sidebar. Through Hive, a new agent appears under Agents like you added it yourself.
Desktop: one agent can ask another and get the reply in its conversation. From CEO, ask CTO — the question shows in CTO's tab, the answer comes back to CEO.
Mobile: tap a notification to open that agent's terminal. The tap connects to the right desktop and lands on that bee.
Mudou
2
Desktop: Grok and Codex routines work like Claude. They use the live bee in the sidebar, not a one-shot run.
Desktop: the routine picker shows the name you see in the sidebar. Hover a routine for run, pause, and edit.
Removido
1
Desktop: the unused Discussions panel is gone. That empty meeting view is no longer in the window.
Corrigido
5
Desktop: on Windows, Claude comes back to the same conversation after a restart. Resume no longer fails looking for a session that was there.
Desktop: asking Grok or Codex from another agent no longer hangs after they already answered.
Desktop: notifications use the bee's name. A Grok bee no longer shows up as "Claude Code", and the same hook no longer fires twice.
Desktop: pasting an image into the terminal works. A screenshot or a copied image from the browser now reaches the agent the same way it does in Chat.
Mobile: notifications group by agent. Each bee's alerts stack in their own thread.
Desktop: attach files from Chat with a + button. It sits next to the mic — pick a file instead of only dropping or pasting it.
Desktop: Windows installs are signed. The publisher is Ebrahim Paula Leite, so Windows can tell who made the installer. SmartScreen may still warn for a while on a new publisher.
Mudou
1
Desktop: a new Chat session opens with an invitation. The Hive mark sits in the empty conversation with “What are we building today?” instead of “no messages yet.”
Removido
1
Desktop: Settings no longer has a Claude session recap switch. Recap already left the sidebar, so that toggle did nothing you could see.
Corrigido
2
Desktop: a new Chat session no longer shows the previous recap. Starting an agent used to keep the last session's summary on an empty conversation.
Desktop: the first Chat message sits above the composer. It used to stick to the top of the pane with a blank stretch underneath.
Desktop: your installed skills show up in the Chat composer. Typing / lists them right under Claude's own commands — user and project skills, each labeled — and picking one runs it in the conversation. A skill that expects input waits for you to finish typing instead of firing bare.
Desktop: images and files you send render as attachments in Chat. A screenshot shows as a thumbnail with your text as the caption — click it to view full size, Esc closes. Other dropped files become compact cards that open in your file manager. An attachment the system has since cleaned up says so, instead of pretending to be clickable.
Mudou
3
Desktop: flipping between Terminal and Chat is instant. The conversation you were reading paints immediately instead of reloading from a blank screen every time.
Desktop: day-long conversations stay light. Chat skips drawing what's scrolled out of view, so long sessions scroll and stream without the whole transcript weighing in.
Desktop: the /resume list opens fast in projects with lots of history. It now reads only the sessions it's actually going to show.
Corrigido
14
Desktop: stopping a message no longer glues it onto the next one. Interrupting the agent and typing again used to send both joined together ("teste" + "opa" arrived as "testeopa") — and could corrupt commands the same way. Every send now carries exactly what you typed. Messages between agents get the same fix.
Desktop: the mode chip follows the agent into plan mode. When Claude decides to plan on its own, the composer kept showing the old mode for the whole turn.
Desktop: a new session starts clean. Stopping a session and starting a fresh one no longer greets you with the previous conversation's answered-choices card.
Desktop: replies show up right after the agent finishes. A reply could take several seconds to appear after the working indicator stopped; it now lands in under a second.
Desktop: /clear no longer leaves the conversation blank. When the fresh session takes longer than expected to appear, the chat recovers on its own instead of sitting empty until the agent speaks again.
Desktop: a moved chat pane keeps updating. Moving a pane between splits — or fast pane changes right after launch — could silently freeze the conversation until you flipped views.
Desktop: Enter with the command palette open no longer picks a resume session underneath it.
Desktop: agents that were running when you quit come back where you left off. After an update they could show as running with an empty "Starting…" pane — a new blank session, not the conversation you had. They now resume that session and show what already printed.
Desktop: Chat keeps /usage in the conversation. It used to yank you into the terminal for Claude's usage screen. The same numbers now land as a card in Chat — once, not a stack of the same cost and duration lines.
Desktop: Chat keeps /resume in the conversation. The list matches Claude's picker — title, when it last ran, branch, and size — without leaving Chat.
Desktop: Chat keeps /clear in the conversation. It starts a fresh session here instead of yanking you into the terminal. The previous one stays in /resume.
Desktop: stopping an agent in Chat stays in Chat. The conversation remains; Start and Resume sit where the composer was, instead of dumping you onto the raw terminal.
Desktop: Esc or a new message closes Chat usage and resume. The usage card and the resume list go away, and you stay in the conversation.
Desktop: Resume after switching chats comes back to the one you switched to. Switching with /resume and then stopping used to send Resume to the previous conversation.
Desktop: a new agent starts as soon as you add it. After you click Add agent, the session boots immediately — no extra play click.
Desktop: HiveTerm opens filling your screen. The window takes the full height and width of the display you launch it on (Dock and the taskbar stay visible). You can still resize it afterwards.
Desktop: copy a plan from Chat. A copy button sits at the end of the written plan, same as on a reply — you get the Markdown, not the rendered text.
Mudou
2
Desktop: Chat tools read as a quiet log. Reads, bash, and edits sit together — read Timer.tsx, edit index.ts +54 −36, bash git status — instead of a stack of fat capsules. Edits start closed; click the numbers for the hunk. A workflow is its title, not a dump of paths. The conversation is a bit wider.
Desktop: click the effort chip to cycle levels. Same gesture as the routine composer. Max uses the same wash as the terminal; Ultracode is the last step, in violet. Model and effort stay locked while the agent is working, so those clicks no longer pile up for later.
Corrigido
7
Desktop: Start starts the agent on the first click. Agents (Claude, Grok, and the rest) no longer sit on "Starting…" until you hit Restart.
Desktop: Chat no longer pastes leftover terminal output into the reply. Collapsed bash stays out of the conversation, option cards keep the full text, and a leftover Planning path no longer appears.
Desktop: you can click anywhere in the composer to type. Empty space next to the model chip or send button no longer steals the cursor — only the real buttons do.
Desktop: Chat shows Claude API errors. An overload or rate limit (and the retry countdown) appears under Working…, in its own card — not only in the terminal.
Desktop: Chat keeps the effort you actually have. The chip follows what Claude kept — after you say No, after a restart, and while Ultracode is running. The "change effort?" question shows up in Chat right away.
Desktop: multi-step questions in Chat stay on the right card. Going Back keeps Feature on Feature, Submit does not vanish, Cancel is no longer junk characters, and after you send, the conversation shows what you picked.
Desktop: finished file reads no longer sit with a blinking amber dot. That light meant "still working" and stayed on after a screenshot or image was already read.
Desktop: HiveTerm is easier on the battery when it's just sitting there. Terminals you aren't looking at no longer keep animating in the background, and Chat stops extra work once the conversation is idle. You should see it less often under "High energy usage" on a Mac.
Desktop: dropping a file or image on Chat attaches it there. It was landing in the terminal underneath, so the agent never got the upload.
Desktop: what you were typing in Chat survives switching agents. Come back and the draft is still there, including files you already attached.
Desktop: files and images sent from Chat actually reach the agent. The chat was sending a placeholder ([Image #1]) instead of the file, so Claude saw the label and nothing to open. Sending also left a ghost copy of your message on screen after the real one arrived. Both are gone.
Desktop: an experimental Chat view for Claude. Turn it on in Settings → Experimental and a running Claude agent can be read as a conversation instead of its terminal — reply from there, attach files, mention project files with @, pick the model and effort, use voice, run slash commands, and stop the agent without leaving the chat. The terminal is always one click away. Claude only for now.
Desktop: answer the agent from the chat. When Claude asks a question — allow a tool, pick the next step of a plan, or fill in a multi-step form — the options appear in the chat. You don't have to jump to the terminal. Permission mode sits on the composer (click or Shift+Tab) and stays in sync with what the terminal is showing. Written plans stay in the chat, including after a restart.
Desktop: copy a reply as Markdown. A copy button sits on every turn, and Cmd+C over a selection copies the Markdown, not the rendered text.
Mudou
2
Desktop: skipped versions update in one jump. If you're a few releases behind, the banner shows where you are and where you'll land — one restart installs the latest. The notes list everything you missed, not only the newest release.
Desktop: after an update, HiveTerm comes back to the front. Restarting for an update no longer leaves the window behind the Dock.
Corrigido
9
Desktop: closing the lid overnight no longer signs you out. Coming back in the morning could show "Reconnecting…" forever or ask you to log in again even though your account was fine. HiveTerm now keeps the session and reconnects on its own.
Desktop: coming back to the app no longer freezes the terminal. Switching away for a while could lock typing, paint a gray empty window, or hitch while a huge backlog flushed in one go. The terminal wakes cleanly and stays responsive.
Desktop: paste from Xcode and Finder works like Terminal.app. Copying an error list from Xcode no longer pastes a screenshot file; dragging a folder from Finder pastes the real path.
Desktop: Chat no longer burns CPU and memory when you aren't using it. Hidden chats don't stay mounted, and the background work parks until you open one.
Desktop: the Add terminal button stays put once you already have a terminal.
Desktop: a session with nothing to resume no longer offers Resume.
Desktop: copying from the chat no longer crashes on macOS.
Desktop: recap text is no longer cut off by the terminal width or split with blank lines.
Desktop: two agents could briefly show each other's conversation. Each chat now stays on its own session.
Desktop: routine chats got a full makeover — watch your agent work, live. Opening a routine now lands in an immersive chat: the agent's answer streams in as it's written, with a line telling you what's happening right now ("Thinking…", "Running npm audit…") and a quiet chip for each tool it uses. Works with Claude, Codex, Grok, and opencode.
Desktop: reply to any of them. Conversations can be continued with Codex, Grok, and opencode routines too — not just Claude. Pick a different model or reasoning effort for a single reply right in the composer (Claude also gets the Ultracode tier), attach files and images with drag & drop, and stop a run mid-flight with one click.
Desktop: routines live inside each project now. A ROTINAS section in the project tree shows every routine with a status badge — pulsing green while it runs, green when the last run succeeded, red when it failed, amber when paused, gray if it never ran. Add one with the "+" row or ⌘⇧R, and right-click for everything else: run now, pause, edit, duplicate to another project, delete.
Desktop: each routine opens in its own tab. Routine chats sit next to your agent and terminal tabs, and each one remembers exactly which conversation you were reading — leave, come back, and you're where you left off. Closing a tab focuses its neighbor.
Desktop: notifications you can click. A routine's result notification opens its chat at the right conversation; failure alerts, agent-finished notices, terminal bells, and agent-sent messages open the exact agent or terminal they came from — switching workspace and project for you if needed. When a routine ends by asking you something, the notification leads with the question and the chat offers one-click reply buttons.
Desktop: routine results reach your phone. Paired phones now get a push when a routine finishes or fails, just like hook notifications.
Mudou
3
Desktop: the separate Routines screen is gone. Everything moved to where you already work: the project tree, the right-click menu, and the chat itself. Templates became one-click starting points inside the create dialog, and creating from a project skips the project picker entirely (a small link brings back multi-project setup when you want it).
Desktop: notifications by default show up even while the app is focused — you can bring back "only when hidden" in Settings. The "Sound" toggle now actually plays the system sound (it previously did nothing).
Desktop: routine notifications stay quiet while you're watching that chat — replying to a routine no longer pings you about the answer you're reading.
Corrigido
5
Desktop: the "Agent notifications" switch now controls what it says. Agent-sent notices ignored the switch entirely; turning it off silently muted a different kind of notification instead.
Desktop: Kimi routines failed to start on current Kimi versions — every run errored immediately. They run again.
Desktop: the start button needed two clicks for a freshly opened agent or command. One click starts it now.
Desktop: a crash-looping process no longer floods you with alerts — failure notifications for the same process are limited to one per minute.
Desktop: notification permission granted after startup is picked up right away — previously the app stayed silent until restarted.
Desktop: one routine, several projects. When creating or editing a routine, pick more than one project — the same task runs in each of them on the same schedule, and each project keeps its own run history. Pause, edit, or delete once and it applies to the whole set.
Desktop: choose the model for each routine. The routine form now has a model picker with the current lineup of each provider by name — Fable 5, Opus 5, Sonnet 5, GPT-5.6, Grok 4.5, Kimi K3 — or type any model id yourself. The list keeps itself up to date: when providers ship new models, they appear in the picker without waiting for an app update.
Desktop: routine instructions from a file. Custom-agent routines can point at a prompt file instead of pasted text — the file is re-read at every run, so editing it updates the routine automatically. Files inside the project travel with it across machines.
Mudou
1
Desktop: routines follow your workspace. The Routines panel now lists only the routines of the active workspace's projects, matching how the rest of the sidebar works. Scheduling is untouched — every routine still runs at its time no matter which workspace is open. The sidebar entry also got a cleaner look, styled like the workspace selector.
Corrigido
1
Desktop: a "trial expired" notice could appear for accounts that were never expired — right after switching accounts or during startup, the app could briefly mistake an active Pro account for an expired trial and flash the upgrade prompt. It now waits until your session is fully loaded before deciding anything.
Desktop: Routines — put your agents on a schedule. A new Routines section in the sidebar lets you pick an agent and a project, write the task, and choose when it runs: daily, weekdays, weekly, or every N hours, in your own timezone. HiveTerm runs it for you and keeps the complete result of every run — restart the app and the history is still there.
Desktop: every run is a conversation you can continue. A run opens as a chat: your task on one side, the agent's answer on the other. When the agent ends with a question — "want me to commit this?" — just reply, with file and image attachments if you need them, and it picks up exactly where it left off.
Desktop: watch and control runs. Follow a run live in its own terminal, stop it mid-run, and get a notification when a routine finishes or fails (configurable per routine).
Desktop: missed schedules are your call. If HiveTerm was closed at the scheduled time, each routine chooses its own policy: run as soon as the app opens, or skip to the next slot — skipped slots show up in the history so you always know what happened.
Desktop: start from a template. Morning briefing, issue triage, dependency audit, weekly changelog draft, test health check, and branch cleanup — one click pre-fills a routine you can tweak.
Routines work with Claude, Codex, Grok, Kimi, OpenCode, and Aider. The Free plan includes 2 routines; Pro is unlimited.
Desktop: unstable internet no longer signs you out — on flaky connections (satellite internet like Starlink, mobile hotspots, spotty Wi-Fi), HiveTerm could suddenly show "Access revoked" and demand a fresh sign-in even though your subscription was fine. A brief drop at exactly the wrong moment made the app mistake the outage for a revoked account. It now rides out the drop, keeps your session, and reconnects on its own — the lock screen is reserved for the real thing.
Desktop: Start begins a fresh session again — on Windows, pressing Start on a stopped agent could pick up the previous conversation instead of starting clean, so there was no way to get a blank slate without Resume doing the same thing. Resume is unchanged and still continues where you left off.
Desktop: you stay signed in — HiveTerm could lock you out with "Access revoked" while your subscription was perfectly active, usually just after it reconnected. Two parts of the app renewed your sign-in at the same moment, and each treated the other's success as its own failure. If you hit this, signing in again is safe and nothing was lost.
Desktop: clone a repository — "Clone a repository…" now sits next to "Open local folder" in the add-project menu and in the command palette. Paste a URL, pick where it goes, and HiveTerm clones it with live progress and opens it as a project. The folder name is filled in for you, cancelling cleans up after itself, and the folder you cloned into last is remembered for next time.
Corrigido
9
Agents start the first time — an agent could sit on "Starting…" with an empty pane while it was actually running perfectly well behind the scenes, and only clicking Restart brought it into view. Whatever it had already printed is now shown.
Editing hive.yml outside the app reloads again — in the installed app the file was not being watched at all, so changes made in an editor were ignored until relaunch. This only ever worked when running from source.
Restarting an agent keeps it an agent — restarting a Queen-spawned sub-agent quietly turned it into a plain command, losing its instructions and environment along the way.
Creating a pull request can't hang forever — if GitHub asked for credentials that never arrived, the app waited indefinitely. It now gives up like every other network operation.
Terminals stop fighting over their own size — switching tabs while the layout changed could resize a terminal twice in one frame, leaving the pane briefly mismatched.
Long sessions stay light — closing terminals released less memory than they should have.
Notifications arrive — push has been silently dropping notifications; they now reach the phone and open correctly.
Fewer permissions requested — the app no longer declares background capabilities it never used.
Purchases on Android — an inapplicable purchase link no longer appears.
Segurança
5
Your settings and agent configs are never replaced by defaults — if HiveTerm could not read a settings or agent-configuration file, it would fall back to defaults and then save those over your real file. It now leaves the file untouched and tells you instead.
Discarding a change stays inside the file you picked — the discard action accepted an empty selection, which resolved to the whole project folder.
File actions stay inside the open project — every command that reads or writes a file now applies the same confinement rule, including for paths reached through a symbolic link.
A lapsed subscription ends phone access immediately — a downgrade left live phone sessions connected until the app happened to refresh, and a subscription cancelled while the desktop was offline was never noticed at all. Voice from the phone is now covered by the same rule.
Only your own dev servers are reachable from the phone — a terminal could add a port to what the phone was allowed to reach; only the commands you defined can.
Mobile: open login links with one tap — when an agent prints a link (like Claude's sign-in page), a chip appears over the terminal to open it in an in-app browser or copy it. No more fighting to select a long URL by hand.
Mobile: arrow keys — the key bar now has ↑ and ↓, so Claude's /resume picker, menus, and shell history finally work from the phone.
Mobile: terminal loading indicator — opening a terminal shows a spinner while it connects instead of a black screen.
Desktop: agents work better together — Claude and Codex now receive built-in guidance on how to talk to each other through HiveTerm, so multi-agent discussions start correctly without extra setup.
Mudou
1
Desktop: working offline no longer locks you out — HiveTerm now tolerates being offline for up to three days before asking you to reconnect. A brief internet hiccup or a service outage never freezes your work mid-task again.
Corrigido
5
Desktop: no more "You're offline" after waking your computer — time asleep was being counted as time offline, so coming back the next day greeted you with a lock screen.
Desktop: the "Retry connection" button now actually retries immediately instead of appearing to do nothing.
Desktop: if your session was signed out remotely, the app now says so and offers to sign in again, instead of showing an endless offline screen.
Desktop: starting Claude on Windows with the phone connected no longer hangs at "Starting…" — the phone's terminal was interfering with the agent's startup handshake.
Desktop: the "Starting…" overlay can no longer get stuck covering a terminal indefinitely.
Longer voice notes — talk to an agent for up to 3 minutes at a time (up from 1), on both the desktop and the phone.
Remove agents from your phone — delete an agent or command from a project in the mobile app, the same way you can on the desktop.
Mudou
2
More voice-to-text every day — the built-in transcription now gives Pro users a full hour a day.
A smoother mobile home — the "My work" summary loads without the screen jumping, projects group into tidy cards you can expand and collapse, and tapping a summary like "Waiting for you" takes you straight to what needs you.
Corrigido
3
Voice notes that can't be transcribed now explain why in plain language instead of showing a raw error.
Starting a stopped agent from the phone shows its terminal right away, at the correct width — no need to back out and reopen.
Swiping a panel closed in the mobile app now works every time.
Preview modern dev stacks on your phone — apps that load their styles and scripts from a separate dev server (like Laravel with Vite) now display correctly in the mobile preview.
Preview Docker Compose apps — a docker compose up project now gets a preview button on the phone, pointed at the right address.
Automatic mobile updates — the app now updates itself in the background, so improvements reach your phone without waiting on an app-store release.
Mudou
2
Connecting to a desktop is much faster — usually a couple of seconds instead of up to fifteen.
Clearer connection status — the app distinguishes connecting for the first time from reconnecting, and explains any failure in plain language.
Corrigido
2
A phone with an expired session no longer gets stuck retrying; it asks you to scan the pairing code again.
Fixed a rare case where signing back in could leave the phone's session unusable.
Instant, clear connection errors — when a desktop can't be reached, the phone tells you why in about a second instead of spinning for half a minute.
Unpairing syncs both ways — removing a phone on the desktop now removes that desktop from the phone right away.
Cancel a connection attempt from the home screen.
Corrigido
3
Fixed the core problem where a desktop showed the wrong online status after sleep or a network change — the connection recovers on its own within seconds now.
Returning to the app refreshes the connection immediately, and the pull-to-refresh spinner no longer spins forever.
Brief network hiccups no longer drop an active session.
Segurança
1
Fixed a sign-in token that could stay valid longer than intended. Tokens were still single-use and tied to your session; the intended time limit is now enforced.
HiveTerm for iOS and Android — a full companion app, not a status widget. Home shows your paired desktops, workspaces, and agents; open any bee and you get a real terminal session streamed from your machine.
Pair in seconds — generate a QR on the desktop, scan with the phone, confirm a short code. One scan signs you in and links that phone to this computer. Unpair anytime from either side.
Push when an agent needs you — leave the desk. When an agent finishes, crashes, or asks for input, your phone gets a notification. Contents stay encrypted end-to-end; only your device can read them.
Works on cellular, not only home Wi‑Fi — the link between phone and desktop is built for the real world (cafés, trains, 5G), not just the same network.
Talk to agents from your pocket — type in the terminal, use the key bar, or hold the mic and speak. Voice uses the same Pro transcription path as the desktop.
Source Control on the phone — see what changed, review diffs, commit and push without opening the laptop.
Preview your local app — open the desktop’s running dev servers (the ones HiveTerm already detects) in a browser on the phone, through the same secure link.
Send photos and files into an agent session from the phone camera roll or file picker.
Start, stop, and resume agents from the phone when you’re not at the machine — including a clear overlay when something has stopped.
Home built for ops — pin the bees you care about, jump back into recent work, see sub-agents under their parent, and which OS each desktop is on.
Install from the desktop — Settings → Mobile → iOS or Android shows a QR that opens the install page on your phone (no multi‑hundred‑MB download dumped on your Mac).
Mudou
3
The phone UI feels like a product, not a prototype — light and dark themes, language support, proper back gestures, smoother sheets, and a first-run welcome that explains pairing.
Cooler and snappier on long sessions — less battery heat from terminals and background work; connects faster when you’re on the go.
Projects stay in sync both ways — reorder projects on the phone or the desktop and the other side follows.
Corrigido
5
Reopening a terminal on the phone after a long agent session no longer shows a screen full of garbage characters. History is cleaned up so colors and UI codes don’t leak as plain text.
The live project list keeps updating even when agents are streaming heavily.
Terminal size follows the phone while you’re attached — no more fighting the desktop pane for control of rows and columns.
Fewer “stuck” states — Start when an agent is stopped, fewer failed reconnects, previews that survive a desktop restart, and cleaner navigation.
Sign-out closes Settings properly; projects don’t get lost when you delete a workspace you weren’t viewing.
Stop actually kills the agent and shows Resume again. After 0.33.0, Stop could paint Stopped with only Start (no Resume) while the process stayed alive: the UI marked stopped optimistically without the session id, and the backend skipped the status emit when the bee was already gone from the map. Stop now awaits kill_bee, returns { killed, resumeSessionId }, always emits stopped with the session id from hooks/store, and tears down the full descendant tree (not just the shell process group) so Claude/Grok children cannot survive as orphans. Simple agent commands also spawn without a $SHELL -c wrapper on Unix so the managed PID is the real CLI.
Source Control panel (VS Code-style, first tab of the right rail): always-visible commit box, per-file staging with Staged / Changes / Merge Conflicts sections, stage-all / unstage-all, and per-file discard. Commit & Push actually pushes now.
Pull, Push and Fetch built in — publishing a new branch sets the upstream automatically, and failures come back as clear messages (authentication needed, merge conflict, diverged branch, network) instead of raw git output. A clickable ↓/↑ sync widget in the header opens the panel.
AI refinement for voice transcripts (Settings → Voice): *Clean* strips filler words and fixes punctuation; *Structured* also reorganizes your ideas into coherent paragraphs and lists. Built into Pro — no API key needed. Formatted transcripts paste into agent inputs as real line structure without auto-submitting.
Restore agents on launch — each agent keeps its own CLI session and resumes that exact conversation when you reopen the app. Supported: Claude, Grok, Codex, Kimi, OpenCode, Antigravity (Settings → Agents; on by default).
Reopen on the last active bee — the app now lands on the bee you were focused on when you quit.
Resume for stopped Grok agents — same overlay button Claude already had.
MCP tools request and respond — one-call ask-and-wait for agent-to-agent conversations.
Structured agent-messaging logs for delivery, identity and drop diagnostics.
Mudou
6
Settings opens as a centered modal (like you'd expect from a desktop app) with a navigation rail — no more full-screen takeover pinned to the tab bar.
Voice transcription is much faster: audio is downsampled and shipped in a single step instead of bouncing through the UI layer, with proper anti-aliasing so accuracy improved too.
Quitting no longer asks for confirmation — running agents are snapshotted for resume and shut down cleanly.
Agent-to-agent delivery is at-least-once: messages are held until they're confirmed on the wire and requeued on any drop, so recipients can no longer miss them. Agents should treat message ids as idempotent.
Spawning an agent with a task now delivers the task once, on the message bus.
Source Control rail got a wider default and a sturdier layout at narrow widths.
Corrigido
7
Agent-to-agent messages no longer vanish when a connection drops mid-delivery — the number one cause of "my sub-agent never got the message".
Trial expiry is handled properly: the sidebar badge now says "Trial expired" instead of a broken "0 days left", and the end-of-trial dialog appears at the moment the trial ends — even offline — with the option to upgrade or continue on Free.
Voice input is on by default for Pro — a legacy setting kept it hidden for accounts created on older versions.
Markdown files opened from Source Control now show the diff, not the rendered preview.
Editing hive.yml no longer silently deselects the bee you're working in.
Quitting no longer wipes agent-resume state, so agents come back running instead of stopped.
Voice transcripts can no longer inject control characters into the terminal.
Workspaces, right in the sidebar. Your workspaces now live as a list above your projects: one click to switch, drag to reorder, right-click to rename or delete, and a "New workspace…" row with its shortcut. Hover a neighbor row to see the cycle shortcut for it.
Move projects between workspaces. Right-click a project → "Move to workspace". Everything keeps running — the project simply shows up in its new home, ready when you switch there.
Switch workspaces from anywhere. Cycle with ⌘⇧] / ⌘⇧[ (remappable in Settings → Hotkeys), or type a workspace's name in the Command Palette (⌘K) to jump straight to it.
Know when an agent needs you in another workspace. If an agent asks for input, finishes, or crashes in a workspace you're not viewing, HiveTerm notifies you even while the app is focused — the notification names the workspace, and an amber dot marks it in the sidebar until you visit.
Split panes across projects. Working on two projects at once? Drag an agent from any project onto a pane and they share one view. Each pane shows its project's name so you always know which repo an agent is acting on, and the split stays reachable from both projects' sidebars.
Mudou
2
Deleting a workspace no longer strands its projects. They move to your first remaining workspace — nothing is deleted, and the confirmation says exactly that.
A pane lives in one split at a time. Adding a pane that already belongs to another split now moves it over instead of duplicating it.
Corrigido
13
The offline screen stops ambushing you. A quick network blip — router reboot, wifi hop, laptop waking from sleep — flashed the full-screen "You're offline" lock instantly. Now it only appears after five continuous minutes offline; short drops show the small reconnecting banner and get out of your way.
Notifications from sub-agents reach you on macOS. An agent calling the notify tool was silently dropped; it now lands like every other notification.
Notification texts no longer show a literal "{workspace}" placeholder.
Token counters stick to the right agent. With two or more agents running in the same folder, usage numbers could attach to the wrong one mid-session.
Crashed agents restart once, not twice. Clicking Restart right after a crash could collide with the automatic restart and bounce the agent twice.
Fast pane closing no longer leaks. Closing terminals quickly (or switching projects mid-spawn) left invisible listeners behind, slowly weighing the app down over long sessions.
Starting many agents at once is steadier on Windows. Simultaneous spawns could slip past the pacing that prevents them from tripping over each other's shared files.
The app opens even with a malformed Claude settings file. A hand-edited ~/.claude/settings.json with unexpected structure could stop HiveTerm from starting.
Voice capture can't freeze the window anymore. A slow microphone driver used to hang the whole UI while the recording started or stopped.
Resized panes keep their proportions. Adding or closing a pane no longer scrambles the sizes you dragged.
Update toast and font picker speak your language. A few stray English strings now follow the app language (English, Português, Español).
No more console flash on Windows when viewing a file from an older revision.
Workspace rows no longer trap a focus ring after switching by keyboard.
Segurança
2
Temporary files for clipboard images and agent-config writes now use unpredictable names, closing a local file-clobbering vector.
Secret redaction now also covers error messages from voice transcription providers.
See what's new before you update. When a new version is ready, a button appears above "Add project" and a modal shows the release notes with a one-click "Restart & update". The download already happened quietly in the background, so the restart is instant.
Mudou
3
Updates no longer install behind your back. HiveTerm checks for and downloads updates in the background, then lets you choose when to restart — no more surprise relaunch mid-task. "Check for updates" in Settings opens the same what's-new modal.
Quieter when idle. Token-usage tracking now wakes on agent output instead of waking on a fixed timer, so the app stops keeping your machine busy when you're not running anything.
Cursor blinks only where you're typing. In a split, only the focused terminal pane blinks its cursor.
Corrigido
1
Copy an agent's output on Windows. Ctrl+C now copies what the agent printed even when the terminal holds no live text selection, instead of doing nothing.
Give a spawned agent its first task up front. When you spawn an agent over MCP, you can now hand it the task to start on right away — it gets to work the moment it's up, instead of sitting idle at its prompt until something nudges it. Set a persona, hand off a task, or both.
Corrigido
1
First hand-off always lands. A just-spawned agent now reliably receives and acts on the first message sent to it, even while it's still booting — so MCP orchestrations stop stalling on a reply that never comes.
Agents with a persona file now start. An agent that only points at a system-prompt file (with no command of its own) launches the configured agent CLI with that prompt, instead of silently doing nothing.
Editing sub-agent instructions no longer breaks the app. Setting a sub-agent's instructions from Project Settings used to corrupt the project config — sometimes leaving an agent that wouldn't start, or a project that wouldn't open at all. Sub-agent instructions now save to their own store, and a single malformed entry can no longer take the whole project offline.
Sub-agents stay under their orchestrator. After a restart, MCP-spawned sub-agents keep their place nested beneath the orchestrator that launched them instead of scattering into a separate list.
Renames stick. Renaming an agent now survives a restart, and turns off auto-rename so the terminal title can't overwrite your chosen name.
Your voice, alive. Hold the mic and a swarm of light blooms across the screen, breathing with your voice as you speak — a big timer front and center, and Esc or your hotkey to stop without reaching for the mouse (Pro).
Markdown, your way. Markdown files open in the editor for quick edits, with a one-click toggle to a rendered preview that reflects what you're typing.
Mudou
2
Files open in an editor, not a viewer. Opening a file now drops you straight into an editable CodeMirror editor — read by simply not typing, and edit and save (⌘S) the moment you want to. In-file search, go-to-line and smooth navigation of huge files all carry over.
Antigravity and Kimi join Discussions. Both can now be picked as discussion participants and orchestrators, right alongside Claude and Codex.
Removido
1
Open-in-external-editor. The built-in editor replaces it — files open directly in HiveTerm, no round-trip to another app.
Corrigido
3
Agents wake each other up. A freshly spawned Claude agent now reliably receives and acts on a message sent right after it boots, so MCP orchestration stops stalling on the first hand-off.
Kimi spawns and runs. Launching Kimi with a task no longer crashes, and turning on skip-permissions now applies its --yolo flag so it doesn't wait on approvals.
Copy on Windows. Selecting text in the terminal and copying it works again on Windows.
Talk to your agents. Hold the mic, say your prompt, and it lands in the terminal as text — no typing, no API key to wire up. Transcription runs on HiveTerm's own service, ready the moment you are (Pro).
Kimi joins the swarm. Kimi Code (Moonshot's K2) is now a first-class agent — spawn it, drop it into a discussion, and let it drive HiveTerm's MCP tools right alongside Claude, Codex and Antigravity. Token usage is tracked and the MCP server connects with a single toggle.
Meet Antigravity. Google retired the Gemini CLI for individual plans, so HiveTerm now ships its successor — Google's Antigravity CLI (agy) — as a first-class agent. Spawn it, orchestrate it, and it drives the Queen's MCP tools like the rest, pre-set to skip approval prompts so it just runs.
Make every shortcut yours. Remap, disable or reset any keyboard shortcut in Settings → Hotkeys. Conflicts are flagged as you bind, and the essentials — tab switching, copy and paste — stay protected.
Mudou
2
Meetings are now Discussions. A clearer name and a redesigned panel: cleaner thread cards and a one-click expand to read any exchange full-screen, with long participant lists and end-of-thread summaries that read at a glance.
A calmer launch. HiveTerm now opens with nothing selected — a clean slate, so you choose where to start instead of landing mid-project.
Removido
1
Gemini CLI. Google shut it down for the free, Pro and Ultra plans — HiveTerm replaces it with the Antigravity CLI (above), which runs the same Gemini models.
Corrigido
2
Free plan: your agent always starts. Relaunching your own agent no longer trips the agent limit — what counts is how many agents you're running, not how many times you've restarted one.
Plan usage tells the truth. The plan panel now shows the real Free limits: 1 project and 2 agents.
Sub-agents that survive a restart. The agents your orchestrators spawn no longer vanish when you quit. They come back in the sidebar, stopped, and one click resumes the exact conversation where it left off — same session, same context (Claude and Codex; Gemini re-runs the task fresh).
Prompt files. Point an agent at a Markdown file as its system prompt and HiveTerm re-reads it on every launch — edit the file, restart the agent, done. Pick one from the Add-agent dialog or Project Settings, and the agent even names itself after the file.
Edit instructions anywhere. Right-click any agent to open its system prompt in a focused editor you can pop out full-screen whenever the box feels cramped.
Personas for spawned agents. Orchestrators can now hand a fresh sub-agent its role at spawn time — loaded as a real system prompt that survives context compaction, instead of being typed into the chat.
One-call agent round-trips. A message can now wait for the reply and return it in the same step, so an orchestrator gets its answer without a separate poll.
Mudou
2
Agents and Commands, separated. Project Settings now keeps your agents and your commands in their own sections — an agent was never really a "command."
The names you (or an orchestrator) give your agents now stick, instead of drifting to the CLI's auto-generated session title.
Corrigido
3
Windows: prompts and paths reach the agent intact. When one agent launches another, the command, the initial prompt and project paths with spaces or special characters are now passed through to the agent intact, instead of getting mangled by the shell.
No more corrupted config. hive.yml is written with a proper YAML serializer, so Windows paths and richer per-agent settings round-trip cleanly.
The sidebar's command-row hover no longer bleeds past the card's rounded corners.
Meetings. A new panel (⌘J) shows your agents talking to each other as clean chat threads — each message with the agent's avatar, replies linked to what they answer, and the final outcome highlighted as a summary card. No more reading raw terminal scrollback to follow a multi-agent exchange.
One-click orchestration. Hit ⌘⇧J (or the + button), pick a template, write the topic, choose the participants. An orchestrator agent spawns the team, runs the conversation and delivers the wrap-up. Six templates: Decision (two agents argue opposing sides), Brainstorm (different perspectives, one synthesis), Review (a defender vs. attackers over your code), Plan (product and engineering lenses merge a plan), Dev squad (splits a feature between agents that build it in parallel in your repo, with exclusive file ownership and a final delivery report), and Custom (write your own brief — roles, rounds and flow are up to you).
Cross-vendor teams. Any of Claude, Codex or Gemini can orchestrate or participate — OpenAI moderating Anthropic vs. Google is now a thing you can do before breakfast.
A real message bus for agents. Under the hood, agents communicate through structured MCP tools (send_message, read_inbox, reply, await_reply) instead of typing into each other's terminals — no noise, no scraping, no lost messages. Replies are correlated, conversations are threaded, and everything works with agents that can't be poked mid-run.
Meeting history that survives. Conversations persist across restarts, load collapsed so you pick what to read, follow new meetings live with auto-scroll, and can be deleted per meeting or cleared entirely.
Agent spawn controls. Agents spawned via MCP can now request a specific model and run with permissions bypassed — so orchestrated teams never stall waiting for an approval click.
Large-file viewer. Huge files open in a virtualized code view with text search, instead of grinding the editor to a halt.
Mudou
3
Splits are now agent-first: the new-split button pairs your agents, and panes can be dragged to reposition or swap directly in the layout.
Empty sidebar sections show inline shortcuts to add an agent, terminal or command right where you'd expect them.
Faster everywhere: meeting messages deliver incrementally instead of reloading the whole history, the file-change watcher on macOS uses a single recursive watch, and several hot paths in the sidebar, palette and git integration were tightened up.
Corrigido
4
Switching workspaces now fully resets the main view — no more seeing the previous workspace's tabs or files.
Agent logos are detected from the actual binary, so an orchestrator that mentions other agents in its prompt no longer wears the wrong badge.
Rearranging split panes no longer garbles the terminal repaint.
Windows/Linux: window controls are back. The macOS-only translucent chrome no longer applies to other platforms, restoring the native minimize/maximize/close buttons that 0.28.0 removed.
Edit files without leaving the terminal. Open a file from the diff or explorer panel and change it right there — a full editor (CodeMirror 6) with syntax highlighting, find, and go-to-line. Save with ⌘S. If the file changed on disk while you were in it, Hive catches the conflict and lets you reload or keep your copy. Saves are atomic and keep your original line endings and BOM, so writing a file never quietly mangles it.
A glass redesign on macOS. Rounder window corners, and a sidebar, diff panel, and status bar that go translucent and pick up the desktop behind them. New in Appearance: a spacious/compact density toggle for how tight you want the rows, plus glass set to on, off, or auto.
Custom window controls (macOS). Hive draws its own close, minimize, and zoom buttons, sitting cleanly under the rounded window instead of jammed into the corner.
A welcome screen when nothing's open. No project loaded? You get a real starting point — open a project, jump to the command palette, or settings — and a list of your recently-opened projects, one click to reopen. Projects that have since moved or been deleted drop off on their own.
Copy a whole file in one click, straight from the viewer toolbar. And ⌘A inside a file now selects the file's text, not the entire window.
Mudou
2
Start a command without it stealing your screen. Hit ▶ on a command or agent in the sidebar and it spins up in the background — your current view stays put. Click the row or tab when you actually want to look at it.
The file viewer handles large files faster, and in-file search feels smoother.
Corrigido
2
Removing your last project now drops you on the welcome screen instead of a blank pane.
Switching a file between viewing and editing no longer shifts the code sideways — the two views line up to the pixel.
Segurança
1
File reads over the IPC bridge are confined to the open project's folder and capped in size, so the app can't be steered into reading arbitrary paths off your disk.
Pins & Notes. Pin any terminal output to a per-project rail and work it as a checklist — for the "found this, not ready to fix it yet" pile. Each pin has an editable title + body, drag to reorder, and done items collapse into their own group. Capture a selection with ⌘⇧P, right-click → Pin selection, or the pill over a selection — and it works inside full-screen TUIs like Claude Code (via OSC 52), not just plain shells. Below the pins, a notes scratchpad autosaves per project. Toggle the panel with ⌘P; expand it to fill the window.
Your agents manage pins & notes over MCP. Eight Hive MCP tools let a running agent list, create, update and check off pins, and read or write notes — scoped to its own project. The agent can pin its plan and tick items off as it finishes them.
Commands in split panes. Command bees are now draggable into the split grid alongside agents and terminals.
Mudou
2
Keyboard shortcuts: the sidebar toggle is now ⌘B (was ⇧⌘B), the branch switcher is ⇧⌘B (was ⌘B), and the diff panel is ⌘Y (was ⇧⌘Y).
New pins land at the top of the list (newest first); drag to reorder.
Corrigido
1
The window opens inside your screen. On small or scaled displays — common on Windows, like 1366×768 or 1080p at 150% — the window no longer opens larger than the screen with its edges cut off. It fits to the screen and centers on first launch, and remembers your size and position between launches.
Performance
1
Editing or capturing a pin no longer re-renders the pins panels of unrelated projects, and an agent's pin edit updates only the project it touched. Also trimmed wasted work in the terminal selection tracker and stopped an internal copy cache from growing unbounded.
Split Panes — your agents, side by side. Split the workspace into a grid and run up to four agents or terminals at once, each live in its own pane. No more alt-tabbing between sessions to find out who's stuck: watch Claude Code plan while Codex implements and your dev server prints logs, all in one window. Drag a bee into a pane to add it, drag the dividers to resize, and split right (⌘D) or down (⇧⌘D) straight from the keyboard.
Named split groups. Save different layouts as named groups — "Frontend", "Review", "Debug" — and switch between them from the sidebar. Fill a split to capacity and the next agent spills into a fresh group automatically, so a click never dead-ends.
Split from anywhere. Right-click any agent to open it beside or below the one you're viewing, or hit "+" on a split to spin up a brand-new agent directly into the grid — in the direction you asked for.
Performance
2
Typing stays instant, even on a huge paste. Your keystrokes and pastes now travel to the agent over a direct binary channel instead of being re-encoded as text along the way. The input lag — and the occasional "Not Responding" freeze on Windows when pasting a large block — are gone.
Lighter and faster everywhere, with the biggest wins on Windows and Linux. Less memory per running agent, quicker sub-agent spawns, a snappier shutdown, and file-watching that no longer trips Linux's watch limits on large monorepos (with a clear heads-up in the diff panel on the rare occasion your system limit is hit). The terminal now adapts its renderer to your hardware, so machines without GPU acceleration — common on Linux — stay smooth instead of stuttering.
Corrigido
1
Sub-agents reliably find their parent. Spawning an agent from another agent over MCP could, after a long session with many connections, fail to recognize which agent requested it. Connections are now released the instant they close, so agent-spawns-agent stays dependable no matter how long you've been working.
Your prompt goes to the agent you actually clicked. Clicking an agent in a project that wasn't focused yet could silently hand your next keystrokes to a different agent — the one the app auto-selected a beat later — so a prompt you meant for "Claude" ended up running in a template generator you never started. Your selection now wins.
Ctrl+A / Ctrl+E / Ctrl+K work in agent prompts on macOS again. The app treated Command and Control as the same key, so the standard readline/Emacs bindings macOS users rely on for cursor movement were being swallowed as app shortcuts (Ctrl+E was opening "Add Agent" instead of jumping to end of line). Command and Control are now distinct everywhere — Control-based editing reaches the agent, Command still drives the app.
Your theme and language really stick now. The 0.25.4 fix flushed preferences on close, but they could still be wiped at the next launch — and on Windows especially — because the saved theme and locale were dropped from the file whenever any other setting was written. Both are now kept in sync on every save, so changing dark mode or your language survives restarts and unrelated setting changes alike.
Performance
1
Large files open without the wait. The file viewer now renders only the lines on screen (the rest stay in the document but skip layout and paint), loads each language's syntax grammar on demand instead of all 37 up front, and runs syntax highlighting off the main thread. Opening a multi-thousand-line file no longer hitches the window, and the very first file you open in a session lands faster.
Your theme and language stay where you put them. Changing dark mode or switching to Portuguese / Spanish and closing the app quickly no longer resets to defaults on the next launch. Every preference is now flushed to disk before the window closes, and the underlying file is written atomically so a crash or sudden quit can never corrupt it.
"This project is already in the workspace" — when it clearly isn't. Adding a project you'd previously removed used to fail with this warning even though the sidebar showed nothing. The leftover workspace entry now gets cleaned up the moment you remove a project, and any stragglers are scrubbed automatically the next time you open the app.
Projects on OneDrive and slow network drives stop disappearing. A project whose hive.yml was a OneDrive placeholder (or sat on a drive that hadn't finished mounting yet) could be silently removed from your workspace on launch. HiveTerm now confirms a path is genuinely gone before scrubbing it, with a tight 200ms cap so a sluggish network can't slow your boot down.
The file viewer no longer "jumps projects" when you close a tab. Closing the last open file in your active project used to surface a file from a different project that happened to be open in the tab strip behind the scenes — the tab bar showed nothing related, and the central view rendered code from a project you weren't in. Closing a file now stays inside the project you're working in, or hands focus back to your last agent when no other files are open.
The Ngrok command stops wearing Grok's avatar. A name-detection rule was matching grok as a substring of ngrok, so xAI's branded icon would appear on the tunnelling CLI in the sidebar. Same fix covers a few other lookalikes (decline, sampler, spider, etc.).
Cleaner agent header. Removed a stray separator dot that lingered next to the token counter after the CPU / memory readouts were retired in 0.25.0.
Mudou
2
The sidebar stops folding away when you stop an agent. Pausing a Claude or Codex session no longer collapses the project section under it — the panel you were actively working in stays put. Auto-behaviour is additive only: starting a bee can still expand a project from elsewhere, but stopping one never hides what you're looking at.
Empty sections start collapsed (but not on a brand-new project). Terminals and Commands sections fold themselves by default when they're empty AND the project already has at least one bee somewhere, so an idle sidebar reads cleaner. Brand-new projects keep every section expanded so the "+ Add" buttons stay visible while you're setting things up. The Agents section is always open — it's the primary thing you'll add first.
Performance
2
Git status and project file listings cache aggressively. A burst of "git changed" events (running git commit, switching branches, a dev server touching a log file) used to fire the same git branch, git status, and full-project file walk over and over within the same second. Each call now checks four .git/ mtimes and reuses the previous result whenever nothing actually advanced — backend wall time on big projects drops by roughly 40%, and the "Not Responding" freezes some Windows users were still seeing on multi-project workspaces are gone.
Cold idle is quieter. Even when no real change has landed, the previous code was paying the full cost of git branch -vv (300-600ms on Windows NTFS) on every watcher tick. The new fingerprint cache turns that into a 1ms cache hit, so a laptop with HiveTerm open in the background stops periodically waking a CPU core for nothing.
No more "Not Responding" on Windows. The freezes users hit when opening multiple projects, switching branches, committing, or pasting large prompts are resolved. Every backend command that used to block the window now runs in the background.
Paste, drop, and voice input stay smooth. Multi-character pastes, dragging files into a terminal, and voice transcriptions no longer pile up dozens of tiny round-trips to the backend — they land in a single batch, which keeps typing responsive even while an agent is streaming output.
Adicionado
1
Debug logging toggle. Settings → Integrations has a new "Enable debug logging" switch. Off by default. When on, HiveTerm records the timing of every backend operation into a rotating log file under ~/.hiveterm/logs/. Lets you attach a precise timeline to a bug report so we can pinpoint freezes without guessing. No personal data, no terminal output.
Performance
2
Boot, git refresh, file picker, and search run off the UI thread. Twenty-six backend commands moved from synchronous to async dispatch. Visible on Windows as a flat UI even with three agents running and a slow git status in flight; visible on macOS as faster boot when many projects auto-open.
Reduced IPC pressure for high-frequency input. Keystrokes, paste payloads, agent prompt injections, and image drops coalesce into a single backend call whenever they arrive in the same frame.
You stay signed in across app restarts. Close HiveTerm, open it tomorrow morning, and you're already in your workspace — no fresh sign-in, no re-paired devices. Sessions persist for 30 days on every platform.
Account page shows your real email. After a cold launch the "Signed in as" line now reflects the account you authenticated with, not a placeholder.
Sessions survive flaky networks. A brief connectivity hiccup on launch (corporate VPN coming up, hotel wifi blip, slow DNS) no longer logs you out. HiveTerm now tells the difference between a transient network error and a real revoke; only the second one ends your session.
Paste images straight into Codex and Gemini. Cmd+V (or Ctrl+V) of a screenshot or clipboard image now works for every agent — Claude Code, Codex, and Gemini all see [Image] in the prompt instead of the literal file path. Drag-and-drop continues to work as before.
Codex and Gemini reliably connect to the Queen MCP on cold boot. Auto-start agents no longer race the MCP server and fail with "MCP startup failed: handshaking with MCP server failed" — they wait for the server to be ready before starting.
Sign in with Google / GitHub on Windows and Linux — completing OAuth in the browser was opening a brand-new (signed-out) HiveTerm window instead of completing the sign-in in the running instance. The deep-link callback now correctly forwards to the running app so the OAuth flow completes in place.
Sub-agent tree view — when an agent uses MCP to spawn helpers, those helpers now nest under it with proper tree connectors in the sidebar. Multi-agent orchestrations are finally legible at a glance: each orchestrator shows its own children, and a fallback "Sub-agents" section still catches anything spawned from outside the bee hierarchy.
Sign in with Google or GitHub — magic links still work, and you can now sign in with your Google or GitHub account from the same screen.
Silent auto-update on restart — no more "Update available" modal. HiveTerm downloads the new version in the background and applies it the next time you restart the app. A small toast confirms the new version with a link to the changelog.
Mudou
4
Windows performance overhaul — the biggest set of Windows-specific perf fixes since v0.8. MCP tool calls, terminal spawning, tab switching, git diff refreshes and project boot are all noticeably snappier. Heavy orchestrator workflows (Codex/Gemini firing many tool calls per second) feel like macOS for the first time.
Cleaner sidebar — removed the CPU and memory readouts. Activity Monitor / Task Manager already do this better, and dropping the polling means less battery drain on laptops.
Cleaner Settings — removed the Privacy tab. Telemetry is anonymous custom events only (no terminal output, no project paths, no personal data), so the toggle wasn't load-bearing.
Account tab shows only desktop seats — admin browser sessions no longer count against your 2-device limit and don't appear in the Active Machines list.
Corrigido
6
No more "License error: not authenticated" on cold boot — agents set to auto-start would race the session restore on app launch; they now wait until the session is ready before spawning.
Tier changes (Pro → Free or Free → Pro) reflect in the app within 10 seconds — the WebSocket connection is more resilient and properly reconnects after token rotation, so admin-side changes propagate live without restarting the app.
Tab switches are faster — removed a redundant render pass on the terminal canvas. Visible on Windows laptops with shared GPUs (200-400ms hitch gone) and marginally smoother on Mac too.
Sign-in dev panel stays open — pasting an OAuth callback URL no longer collapses the accordion on every keystroke.
Sub-agents spawned by sub-agents now nest correctly — fixed an MCP edge case where literal $HIVETERM_BEE_ID shell templates were treated as a parent ID.
Codex agent has a Yolo flag suggestion — --dangerously-bypass-approvals-and-sandbox is now a one-click chip in the Add Agent and Edit Agent modals.
Removido
3
Update modal and "update available" sidebar pill — superseded by silent auto-update + post-restart toast.
CPU/memory monitoring (resource_monitor and the related sidebar badges).
Grok CLI is a first-class agent — xAI's new coding CLI is preconfigured in Settings → Agents, in the Add agent modal, and in the sidebar's right-click → Add agent menu, complete with its --always-approve yolo flag and a proper avatar.
Keyboard hints in the sidebar — the "Add agent", "Add terminal" and "Add command" buttons now display their shortcuts (⌘E, ⌘T, ⌘D) right beside the label, so the muscle memory is one glance away.
Mudou
5
Commit and Commit & Push hand work to your agent — clicking either button while an agent is running in the project now submits the request straight to the agent (no more text sitting unsent in the prompt). The agent picks the conventional commit message and pushes when asked; if no agent is running, the manual commit form still opens as before.
Pro Trial reads as unlimited — the Account tab no longer shows free-tier usage bars during your trial. Backend already grants Pro-equivalent limits while trialling; the UI now matches reality.
Sidebar stays quiet under normal load — CPU and memory readouts on projects and processes only appear when usage actually matters (project: 25% CPU / 1 GB, process: 30% CPU / 500 MB). Flip back to "Always" in Settings → Sidebar if you prefer the live numbers.
Auto-start defaults to off — adding a new agent or command no longer pre-checks "Auto-start when project starts". One less surprise next time you open the project.
Roomier Add agent modal — the agent quick-pick row fits all seven presets cleanly on a single line instead of wrapping or scrolling sideways.
Performance
2
Instant startup on any network — opening six projects at once no longer freezes the app when GitHub is slow or unreachable. The diff panel boots without waiting on remote services, so the UI stays responsive even offline.
No more "Not Responding" with multiple projects — the worst-case startup time dropped from minutes to milliseconds for users on restrictive networks (corporate proxies, VPNs, regional GitHub routing issues).
Mudou
1
Diff panel focuses on local changes — the panel now shows your working-tree changes, branch status, commits ahead/behind, and file tree without depending on any external service. Commit, Push and Create PR buttons keep working exactly as before.
Resume any Claude session in one click — when a Claude bee stops, the new Resume button next to Start respawns it with claude --resume <id>, so the agent picks up exactly where it left off. No copy-paste, no command memorisation.
Five premium monospace fonts, bundled — Iosevka (default), JetBrains Mono, Fira Code, Geist Mono, and the existing system stack ship with the app. Pick your favourite in Settings → Terminal; nothing to install, ligatures included where the typeface supports them, fully offline.
Monospace-only font picker — when you bring your own font, the picker now hides proportional faces by default so you can't accidentally turn the terminal into Times New Roman.
Ghostty-grade terminal capabilities — opt into bundled xterm-ghostty terminfo (Settings → Terminal) to unlock richer colour and graphics support in modern TUIs.
Adjustable letter spacing — fine-tune terminal density from Settings → Terminal; the new default of 1px gives JetBrains Mono and Iosevka the breathing room they were designed for.
Performance
2
30× faster bee spawn on macOS — the PTY pipe now bypasses Tauri's IPC channel through a loopback WebSocket, eliminating the ~23 second head-of-line stall that piled up when more than one bee mounted at once. Spawning ten Claude bees in parallel is now near-instant.
WebGL terminal renderer, smartly — the GPU-accelerated renderer is attached only to the visible pane, so opening multiple projects with several agents each never trips Chrome's WebGL context cap and never repaints garbage glyphs after a tab switch.
Corrigido
7
Token counts pinned to the right Claude bee — running two or three Claude bees in the same project no longer mixes up their token counters. Pairing is now deterministic via Claude's SessionStart hook (zero timestamp guessing), with an mtime fallback for any bee whose hook hasn't fired yet.
"Open Changes panel by default" setting is honoured everywhere — every reader of the per-project diff state now respects your global preference, so toggling it off finally keeps the panel closed.
Settings tab stays pinned — clicking a bee or command while Settings is open now just hides the panel; the tab stays in the tab bar so one click brings you back. Only the X (or Esc) tears the tab down.
Auto-rename choice persists across restarts — the Add Agent modal now writes auto_rename into hive.yml, so an agent you opted in to keeps renaming itself on the next app boot.
Stopped agents always show a clickable Start button — the stopped overlay now sits above the diff/changes layer, fixing the case where the button rendered behind another panel and ignored clicks.
Mascot no longer floats mid-pane after Vite HMR or restart — fonts are loaded before xterm measures the cell, the WebGL atlas is flushed when a pane becomes visible, and restarts mirror the manual Stop → Start sequence so the mascot lands where it should.
No more stray WebSocket connection errors in the console — silenced the cosmetic warning that fired every time a bee restarted or a hidden pane attached.
Mudou
3
Default terminal font is Iosevka — narrow, distinctive, designed for code. Pick another in Settings → Terminal at any time; your system fonts are still available.
Diff panel is closed by default for new projects — the global toggle in Settings → Sidebar lets you flip the default; per-project overrides still win.
Stop Claude hook ships out of the box — installed silently on first boot; powers both the Resume button and the per-bee token counter without needing a manual setup step.
Segurança
1
CSP whitelist for loopback WebSocket — production builds now permit ws://127.0.0.1:* and ws://localhost:* under connect-src, so the new PTY transport works in shipped binaries (not just dev builds).
Windows: Claude Code, npm, Node and other agent bees now spawn reliably — replaced the PTY layer with the upstream portable-pty-psmux crate, eliminating the intermittent crash that left bees stuck in the Crashed state immediately after launch.
Corrigido
2
Crashed bees no longer leave a blank pane — selecting a Crashed bee now shows the agent's exit code and the last 30 lines of output it produced before exiting, plus a prominent Restart button. Previously the pane stayed empty and the only signal was a tiny red status dot in the sidebar.
Auto-starting multiple Claude bees no longer races — when a project auto-started several agents at once on Windows, some would silently crash because the underlying CLI's lockfiles, AppData ACLs, and shell wrapper chain raced on shared resources. Spawn dispatch is now throttled to one bee every 250 ms, removing the concurrent access without any user-visible delay.
Corrigido
2
Windows: claude, npm, node and other npm-installed CLIs now spawn cleanly — the leaner Windows spawn path introduced in v0.22.0 was bypassing cmd.exe for commands without shell metacharacters, which caused npm-style shims (claude, npm, node, python, cargo...) to fail with os error 193: %1 is not a valid Win32 application. CreateProcessW doesn't honour PATHEXT and grabbed the extensionless Bash shim instead of the .cmd sibling. Commands without an executable extension are now routed through cmd.exe again so Windows' PATHEXT lookup picks the right file. Programs given by absolute path with .exe/.bat/.cmd/.com keep the fast path.
Windows: duplicate "Help" menu — the v0.22.1 menubar showed two "Help" entries because the new "Check for updates…" command appended a fresh submenu instead of merging into the default Help that Tauri already creates on Windows and Linux. The command now lives inside the existing Help menu.
Check for updates, anytime — a new menu entry (HiveTerm › Check for updates… on macOS, Help › Check for updates… on Windows and Linux) lets you pull the latest release without restarting the app.
Jump straight to the change — clicking a file in the Changes panel now scrolls the viewer directly to the first modified line, so you stop hunting through long files to find what's new.
See every file, just like your editor — the Explorer now mirrors the VSCode behavior and surfaces files listed in .gitignore (.env, lockfiles, generated docs). Heavy directories like node_modules, target, and dist stay hidden so the tree remains usable.
Mudou
1
Cleaner button labels — the prominent action in the Diff panel is now "Commit & Push" across English, Portuguese, and Spanish.
Performance
6
Lower CPU on every platform — the resource monitor now aggregates CPU and memory across each bee's full process tree, so wrappers like cmd.exe /C and shell pipelines report the work their children are actually doing.
CPU metrics that work on Windows — fixed a baseline-sampling bug where Windows showed a flat 0% for every bee. Per-bee and per-project CPU readings are now accurate on every OS.
Snappier tab switching with chatty agents — the hidden-pane PTY buffer now flushes in 64 KB chunks instead of one 2 MB write, eliminating the brief UI freeze that could appear when switching to a long-running Claude or Codex tab.
Stable memory under heavy git activity — the per-project file watcher now applies back-pressure to its event queue, preventing memory growth during large git checkout or npm install bursts on Windows.
Cleaner file picker on Windows — the Explorer's always-skip list (node_modules, Build, Vendor, etc.) now matches case-insensitively, matching how NTFS treats those directory names.
Self-healing token tracking — when an agent CLI rotates or removes its session log, the token watcher now re-discovers the new file on the very next tick instead of going silent.
Corrigido
4
The "0 bees but limit reached" mystery — the Free-tier per-project limit was incorrectly counting bees from every project in the workspace, so users with multiple projects could be blocked from launching a fresh bee even when the target project was empty. Counting now correctly scopes to the project being launched.
Commands with paths that contain spaces — running node "C:\Program Files\app\index.js" (and any other command with quoted arguments) no longer mangles the arguments before reaching the binary. Affects every platform; most visible on Windows.
No more phantom "crashed" badges on Windows — when Windows recycled a PID inside the auto-restart window, the previous bee's exit could be wrongly applied to a freshly started bee. The watcher now identifies a spawn by (pid, started_at), eliminating the false positive.
"You're up to date" dialog dismisses on click — the OK button on the manual update check now closes the dialog reliably.
Agent instructions library — pick a system prompt from a curated set of dev-focused templates (Software Engineer, Software Architect, Front-End Developer, Back-End Developer, DevOps Specialist, Code Reviewer, Debugger, Refactoring Specialist, Test Author, Technical Writer, Product Developer, UI/UX Designer, Technical Product Manager) or write your own. Instructions are auto-loaded when the bee starts — natively via --append-system-prompt for Claude Code and --system-instruction for Gemini, and via a quick paste-into-prompt for every other CLI. Picking a template also names the bee after the role.
Per-project share toggle — choose whether each project's agent instructions stay private to you (default) or get committed to the repo so the whole team uses the same prompts. The choice is sticky across reloads.
Drag-and-drop files into any terminal — drag any file or folder from the Changes panel or Explorer into a terminal and the path lands in the agent's input as @src/file.ts for Claude / Gemini (native auto-resolve) or as the absolute path for Codex and other CLIs, so the agent opens the file directly instead of grepping for it.
Find in Files — VSCode-style multi-line content search across the whole project, with case-sensitive, whole-word, and regex toggles. Results group by file, expand inline, jump to the exact line + column.
File viewer with syntax highlighting — open any file in the project from the Explorer or a search result. Powered by Shiki, theme-aware, with line numbers, search-match highlighting, and Markdown rendering for .md files.
Editor & terminal font customization — pick from every monospace font installed on your system in Settings → Appearance. Live preview while you scroll the list.
Settings as a tab — Settings now opens as a regular tab in the tab bar instead of a modal overlay, so you can keep it open while you work and switch back to a terminal without losing your place.
Command palette (Cmd+K) — fuzzy-find any file in the project or jump to any action.
Auto-responsive panels — the sidebar and Changes panel collapse automatically when the window is below 600px wide, and restore when there's room again.
Multi-line inline diff — the file viewer now highlights changed *ranges*, not just whole lines, so you can see exactly which characters moved.
Per-agent CLI flag presets — the Add Agent modal now offers correct flag chips per CLI (e.g. Codex gets --dangerously-bypass-approvals-and-sandbox, -s workspace-write, etc. instead of the old --full-auto that the CLI no longer accepts).
Mudou
4
Sidebar tree connectors removed — cleaner, denser sidebar without the dotted ASCII tree lines.
Project Settings layout — directory, config file, gitignore toggle, share-agent-instructions toggle, and command count are now grouped as a clean "Overview" card.
Add Agent modal — collapsible Instructions section, searchable ComboPicker for templates, hidden Name field for agents (auto-generated and adopted from the agent's session title via Auto-rename).
Terminal link provider — clickable links now recognise tilde-prefixed paths (~/Projetos/...) and resolve them against your home directory before opening.
Removido
1
Marketing-flavored templates — Marketing Strategist, Product Marketing Manager, SEO Specialist, Product Designer, and the generic QA Engineer template were dropped to keep the picker focused on the people who use HiveTerm: developers building and maintaining software.
Corrigido
5
Auto-start agents finally auto-starting — auto_start: true on an agent in hive.yml now actually starts the agent on boot. Previously the logic excluded agent-type bees from the auto-start pass.
Paste-and-submit reliability on TUI agents — instructions pasted into Ink-based CLIs (Claude, Gemini, etc.) now reliably submit on first try.
Find in Files preserving indentation — multi-line patterns that depend on leading whitespace now match correctly.
File viewer respecting the chosen font — the browser default was overriding the picked monospace font inside the viewer's code blocks.
Settings tab no longer disappears when opening a file — opening a file from the Explorer keeps Settings pinned in the tab bar instead of closing it.
Workspace loading splash — animated hive logo with a "Loading workspace… (X of Y projects)" indicator while project configs are parsed at boot. No more staring at a blank sidebar when restoring a workspace with many projects
Corrigido
4
Bees not spawning in production builds — spawn_bee was hanging on the license RwLock because phone_home self-deadlocked: a Rust temporary lifetime kept a read guard alive across an assignment that also reached for a write guard on the same lock. The same-thread reader-then-writer pattern blocked forever. The phone-home update now snapshots the read state into local variables before taking the write lock
Spurious downgrade to Free tier on heartbeat — the phone-home heartbeat was overwriting an active Trial/Pro session whenever the server happened to return tier: "free", even though the local signed token (with its server-issued expires_at) was still valid. Trial/Pro expiration is now determined locally from the token. Server revocation still applies via the explicit revoked: true branch
Tauri IPC blocked by Content-Security-Policy in release builds — the production CSP was missing connect-src entries for the Tauri IPC scheme (ipc: / http://ipc.localhost) and the HiveTerm API endpoints, so every invoke() failed silently in release. Dev builds appeared to work because Tauri's dev profile relaxes CSP. The CSP now allows the IPC scheme and the *.hiveterm.com API hosts explicitly
spawn_bee, kill_bee, restart_bee blocking the UI thread — sync Tauri commands run on the webview/main thread, which meant the spawn_pty fork+exec path froze the window when several bees came up at boot. The commands are now declared with #[tauri::command(async)] so Tauri dispatches them on the async runtime while keeping the synchronous signature (and the Channel<Vec<u8>> PTY stream) intact
Bees stuck on the "Stopped" overlay after clicking Start — clicking Start (or any restart action) on a stopped bee fired a kill_bee first, which always emitted a stale bee:status:stopped event. The event reached the frontend after the spawn had already set status = "running" and reverted the UI to the idle overlay, even though the process was up and the PTY was streaming. The frontend no longer sends a redundant kill before each spawn (the backend already tears down the previous process atomically), and kill_bee only emits a stopped event when it actually killed something
UI freeze on multi-project boot — opening a workspace with several projects no longer hangs the window for 30s+. The bee spawn pipeline (spawn_bee / kill_bee / restart_bee) was running on the webview/main thread and serializing every fork+exec through it; it now runs on a background pool so the UI stays responsive while bees come up
Boot-time fork+exec storm — agents listed in hive.yml no longer auto-spawn when the workspace loads. They start the first time you open their tab. Server-style command bees marked auto_start: true still launch eagerly as before
14-day Pro trial — new installs and existing free-tier users can start a 14-day Pro trial directly from the sidebar. No credit card required
Anonymous usage analytics — opt-out toggle in Settings → Privacy. Helps us understand where the product needs work. No PII, no project paths, no terminal output
Trial-expired project picker — if a trial ends with more projects open than the Free tier allows, a blocking modal lets you choose which to keep (or upgrade to Pro to keep them all)
Welcome to Pro modal — short celebration after activating a Pro license that recaps what's unlocked
"Auto-rename" toggle in the Add agent modal — uncheck to lock the agent name when adding it; default keeps the existing behavior of adopting the agent's session title
System theme as the default appearance — fresh installs follow the OS preference and update live when the OS toggles dark ↔ light. Manually picked themes still persist
Pro education accordion in the sidebar — collapsible bullets summarising what Pro unlocks; the whole row is clickable
Mudou
4
Free tier limits adjusted — up to 2 projects, 3 bees per project, 6 bees total, 1 MCP sub-agent. The new 14-day Pro trial covers anyone who needs more while evaluating
Sidebar license footer is a single row (status + Upgrade button + chevron) instead of multiple stacked cards
Localized pricing in the Upgrade and Trial-expired modals — both modals now show your region's price via the existing /api/pricing endpoint instead of hard-coded values
PR status badges (Merged / Open / Closed / Draft / Pending) now use legible colors under the light theme
Corrigido
7
Windows Ctrl+V double-paste — clipboard contents now reach the PTY exactly once. Please re-test on a Windows machine and report back if it still misbehaves
File viewer "Code" mode showing stale content after an agent edits the file — re-opening a file always shows the current contents on disk
Focus lost when closing a file tab — the previously-active bee is restored when the last file tab closes
Setup project modal closing on drag-released-outside — selecting text inside the modal and releasing the mouse beyond its edge no longer cancels the setup
Trial users hitting the Free-tier paywall when spawning bees or sub-agents
First-launch race condition that could auto-start a trial for an existing user
White screen the first time the project setup modal finished detecting the project stack
Segurança
2
Trial token integrity — trial tokens are signed with the same Ed25519 keypair already used for Pro licenses; the private key never leaves the server. Tokens are bound to the issuing machine and rejected on any other
Telemetry payloads are explicitly allow-listed — no clipboard, no file contents, no terminal output ever leave the device
Material Icon Theme — file/folder icons in the FileTree, Changes list, and TabBar now use the open-source Material Icon Theme (1238 SVGs, MIT). Resolution cascades through filename → extension → language id (via Shiki's lang map) so *.ts, *.php, *.rs, package.json, Dockerfile, src/, etc. all light up correctly. Icons copied to public/file-icons/ by scripts/copy-file-icons.mjs (predev / prebuild / postinstall)
"Open file" picker in the All files tab — sticky search input at the top with fuzzy match, keyboard navigation (↑/↓/Enter/Esc), basename score weighted 1.5× over path score. Backed by a new list_project_files Rust command that walks the project respecting .gitignore (max 50k files, skips .git/.DS_Store)
Reveal active file in FileTree — opening a file in the viewer auto-expands every ancestor directory and scrolls the row into view. New revealPath action on fileTreeStore handles the lazy load top-down so rows fill in from the root
Active file highlight in lateral — DiffFileList (Changes tab) and FileTree (All files tab) both highlight the currently active file viewer tab. Closing or switching tabs clears immediately
TabBar auto-scroll — opening a new file or switching tabs scrolls the tab strip so the active tab is always visible
Cmd+B branch switcher — toggle the branch dropdown from anywhere with ⌘B. Chevron rotates 180° when open, dispatched via a window CustomEvent so dropdown state stays local to the header
Commit footer link — the last-commit row at the bottom of the diff panel is now a button that opens <remote>/commit/<sha> in the browser when the project has an HTTPS remote. Falls back to the previous read-only display for SSH-only / private remotes. Remote URL cached per-project in a new lib/remoteUrl.ts helper
Shared fuzzy match library — src/lib/fuzzy.ts centralizes the scoring used by the Command Palette and the new file picker
Mudou
4
Toolbar height alignment — FileViewer toolbar now uses h-9 to match the TabBar / CenterHeader / DiffPanel headers, so the action icons line up with the search input on the right
Open in editor passes Cursor explicitly to the backend so the picker doesn't fall through to $EDITOR (often vim/nano, which silently spawned a hidden TTY)
Removed close (×) button from the FileViewer toolbar — the per-tab close on the tab strip already covers it
Website /changelog — removed the GitHub Releases link since the repo is private
Corrigido
3
Syntax highlighting + line numbers in code viewer — Shiki was failing silently because Tauri's strict CSP (default-src 'self') blocks the default Oniguruma WASM engine. Switched to createJavaScriptRegexEngine() (pure JS, no WASM). Highlighting and the gutter line numbers both come back; PHP, TS, Rust, etc. all render correctly
Project settings overlay blocking file tab clicks — clicking a file tab while the project settings panel was open looked like a no-op because the overlay sat on top. fileViewerStore.clearActiveBee() now also clears showProjectSettings, so activating a file tab dismisses the overlay
Sticky search header in the FileTree — removed the parent's py-1 and the sticky's mb-1 so tree rows no longer bleed through the gap above/below the search bar when scrolled
Multi-project performance — useGitWatcher now reconciles git watchers incrementally instead of tearing down and re-spawning every watcher whenever a project is added, removed, or its config updates. With 9+ projects open, adding/switching projects no longer fans out 27+ git IPC calls per change. Watcher subscription uses a stable id+path signature so unrelated project mutations don't re-run the effect
Visibility refresh scoped to active project — when the app regains focus, useGitWatcher only refreshes the active project's diff state. Background projects keep their watcher running and pick up changes via push events. Drops focus-return cost from 3*N IPC calls to 1
Persistent Diff Panel per project — DiffPanel now stays mounted per project (like terminals do), hidden via CSS display toggling instead of unmount. Switching projects preserves commit-message draft, scroll position, file tree expansion, and active tab. Same persistence for the bottom-sheet variant on narrow windows
Booting overlay in TerminalPane — when a bee spawn or restart begins, a small "Iniciando…" overlay with a spinner is rendered over the terminal until the first PTY byte arrives. Bees attached to an already-running process with buffered output skip the overlay. Spawn errors clear the overlay so the idle state is reachable
Auto-collapse projects with no running bee — projects with bees configured but none running default to collapsed in the sidebar. Empty (no-bee) projects stay expanded so the user sees the empty state. Once the user clicks the chevron, the choice is persisted in settingsStore.projectCollapsedByProject and survives app restarts; the persisted preference always overrides the auto rule
Tighter sidebar tree connectors — bee/section tree elbow indent reduced from 16px (w-4) to 10px (w-2.5) for a more compact hierarchy
Visible kbd on diff panel toggle — the right-edge expand/collapse button shows the shortcut as a kbd badge inline (matching the GitHub Open button style); shortcut removed from tooltip since it's now visible
Corrigido
1
Removed border-bottom from project settings header — the sticky Start all header no longer renders a divider line under it
Right-side Diff Panel — git workspace pane per project: branch + ahead/behind, PR card (state-tinted background — purple Merged, green Open, red Closed/Conflicts, amber Pending, zinc Draft), file list with +/- numstat, two tabs All files / Changes N. Auto-opens by default for every active project; user choice persisted per project. Default width 320px (drag-resize 280–720). Bottom-sheet variant on windows narrower than 1200px
Tab strip in Diff Panel — All files browses the full project tree (lazy read_dir Rust command, respects .gitignore via the ignore crate, skips .git / node_modules / target / dist / .DS_Store and friends). Changes N shows the modified file list with badge counter
Read-only File Viewer — click any file (in either tab) to open it as a tab in the center column. Shiki syntax highlighting (lazy-loaded singleton, ~37 grammars), Markdown render with raw/preview toggle (react-markdown + GFM + sanitize), image preview via convertFileSrc, unified diff render. Cmd+F search overlay with match highlighting + prev/next navigation
Three-column layout — TitleBar removed; each column owns its own h-9 top bar. Sidebar header reserves space for macOS traffic lights when expanded. Center column hosts breadcrumb + workspace folder pill + branch dropdown + Revisar (per-bee changes) + GitHub link. Diff panel column hosts PR chip + state pill + Commit / Push / Create PR buttons in a single bar that takes the PR's tone color
Commit / Push / Create PR — three direct git actions in the diff panel header. Inline forms (textarea + Cancel + Submit) instead of modals. Push only renders when ahead > 0 and an upstream exists, with a counter chip. Create PR is hidden on protected branches (main / master / trunk / develop / development) and shows an "uncommitted changes won't be in this PR" warning when the working tree is dirty. Both forms gain an Ask {agent} button when an agent bee is running for the same project — clicking sends a prompt to its PTY (commit all changes with a descriptive message / create a pull request for the current changes) and closes the form
Discard changes per file — hover-only undo button on each row in the Changes tab, with a confirm dialog before reverting. Tracked files restore via git checkout HEAD -- <path>; untracked files are deleted from disk. Path-traversal-safe
Rust file watcher (push-based, zero polling) — replaces the 3s setInterval refresh with a notify recursive watcher per project. Ignores .git/objects, .git/lfs, .git/logs, *.lock files, node_modules, build dirs, agent state dirs (.claude, .codex, .cursor, .gemini, etc.) — and uses an allowlist for .git/ so only HEAD / index / refs/heads / refs/remotes flow through. Debounced 250ms. Frontend useGitWatcher listens for git:changed and refreshes the diff store, file tree, and CenterHeader branch info
Sidebar diff stats — per-project insertion/deletion totals (+2175 −70) shown next to the project header in the sidebar
Sidebar footer expanded — Update pill, Free-tier usage + Upgrade CTA, Add project, theme + settings now live in the footer instead of a global title bar
Mudou
9
Lazy per-file diff — get_git_changes no longer returns the full diff blob (the heaviest single git subprocess in the refresh path). New get_file_diff(dir, path) is invoked only when a file is opened. Branch-switch and refresh latency drop ~3-5x on big repos
Branch-aware PR cache — PR cache key now includes the current branch, so a branch switch doesn't return the previous branch's stale PR
PR fallback on protected branches — when on main / master / develop / etc. and gh pr view returns no PR, fall back to gh pr list --base BRANCH --state merged --limit 1 so the most recently merged PR shows on the base branch
Project switch is instant — CenterHeader reads branch / dirty / upstream from useDiffPanelStore.branchInfo (already kept fresh by the watcher) instead of fetching get_git_branch per switch. Remote URL cached per project at module level
TabBar height aligned to h-9 — matches sidebar / center / diff column headers across all three columns
Diff panel opens by default — diffPanelOpenByProject[id] !== false (treats undefined as open). Closed only when the user explicitly toggles it off; the choice persists per project
Header height harmonization — every column header strip uses h-9 so the three columns align across a single horizontal grid line
Shortcuts remap — Cmd+Shift+D now toggles the Diff Panel (was the per-bee Changes panel). Cmd+Shift+B toggles the left sidebar. Cmd+\ kept as an alias
Settings → Hotkeys updated with Toggle diff panel row and the new Cmd+Shift+B for sidebar toggle
Corrigido
1
Windows: Enter key submits in Claude Code / Ink agents — process_manager::write_pty no longer translates \r on Windows. PSEUDOCONSOLE_PASSTHROUGH_MODE (Windows 11 22H2+) forwards the input pipe untouched, so Ink's keypress parser sees \r and treats it as Enter the same way it does on Mac/Linux. Earlier translations to \r\n and to \n both landed as newline insertions in multi-line input. PowerShell, cmd, and other cooked-mode shells continue to work because they accept either CR or LF as Enter
Removido
2
ChangesPanel (per-bee fullscreen modal) is gone — the Diff Panel covers the same data with a better UI. The old Cmd+Shift+D binding has been repurposed for the Diff Panel toggle. The legacy Revisar button in the per-bee status bar has been removed
TitleBar component deleted; its responsibilities moved into the per-column headers
Windows: Claude Code / Codex / Gemini raw-mode TUI agents no longer crash on first input — the upstream portable-pty-psmux 0.9.3 enables PSEUDOCONSOLE_WIN32_INPUT_MODE unconditionally, which makes ConPTY re-encode every keystroke as the Win32 Input Mode escape sequence (ESC [ Vk;Sc;Uc;Kd;Cs;Rc _) toward the child. Cooked-mode shells (cmd, PowerShell) decode this fine, but Ink/readline-based TUIs read stdin directly and treat it as garbage — Claude Code's first-run theme picker (v2.1.126) and any subsequent arrow-key/Enter/Esc input would crash the agent. v0.16.4 tried to compensate from the host side by encoding lone \r/\x1b as Win32 Input Mode but PSEUDOCONSOLE_PASSTHROUGH_MODE (Win11 22H2+) forwards bytes raw, so PowerShell saw the literal escape and broke; that change was reverted in v0.16.5. Real fix: vendor the crate as portable-pty-hiveterm under src-tauri/vendor-portable-pty/ and drop the WIN32_INPUT_MODE flag from base_flags in PsuedoCon::new and new_without_passthrough. ConPTY falls back to plain VT input — both raw-mode TUIs and cooked shells handle that. Escape hatch: set HIVETERM_WIN32_INPUT_MODE=1 to opt back in for editors that benefit from the richer key info (vim modifier keys, IME composition). The local normalize_crlf_for_conpty translation in process_manager::write_pty is kept as a defensive no-op for non-WIN32-Input ConPTY (it only triggers on lone \r)
Per-bee recap line in the sidebar — every agent bee can now show a one-line summary of its most recent turn under its name in the project tree. The recap_watcher background thread polls each agent bee's PTY ring buffer (last 16KB, every 400ms) and scrapes Claude Code's native * recap: marker (emitted by Claude when its awaySummaryEnabled setting is on) and the multi-line └ <body> block produced by the /recap slash command. Continuation/wrapped lines are joined into a single sentence; CR-overwrites from TUI redraws are split apart so spinner status doesn't pollute the captured text. The recap is pushed back to the frontend via set_recap on ProcessManager and the new bee:recap Tauri event; long recaps wrap in a full-width row that sits below the bee row so hover-state on the metrics column never reflows the wrap. New REST endpoint POST /recap { beeId | cwd, text } lets external integrations push a recap directly to Queen
Sidebar recap visibility + Claude session recap toggle — Settings → Sidebar adds a "Agent recaps" toggle (default on) that controls only the sidebar render. Settings → Agents → Recap section adds a "Claude Code session recap" toggle that mirrors Claude's own /config → Session recap switch by writing/removing the awaySummaryEnabled key in ~/.claude/settings.json. On open, the panel calls get_claude_session_recap to sync the UI with whatever the user has configured outside Hive — so flipping it in Claude /config and flipping it in Hive stay consistent. Both toggles persist and i18n labels ship in EN/PT/ES
Bottom-of-terminal breathing room — a fixed-height spacer (h-2) sits between the xterm container and the status bar so the last line of TUI output (Codex's input box, Claude's Cooked for Xs, plain shell prompts) doesn't crash into the status bar. The spacer is outside the xterm container so the FitAddon row math is untouched — earlier attempts to add bottom padding inside the xterm parent caused descenders of the last visible row to be clipped. Status bar height also bumped from py-1 to py-2 for visibility
Mudou
4
strip_ansi translates CSI <n> C (cursor forward) into N literal spaces — Claude Code's TUI renderer emits inter-word spacing as cursor-move escapes rather than U+0020 chars (e.g. Opus\x1b[1C4.7\x1b[1C(1M\x1b[1Ccontext)). The previous stripper consumed the escape and dropped the visual gap, producing Opus4.7(1Mcontext) in the captured recap. The rewrite captures the CSI parameter, dispatches via emit_csi_visual based on the final byte, and emits N spaces for C / a (HPR). Other CSIs (SGR m, cursor up/down, etc.) are still stripped silently — they have no visible content to preserve
ProcessInfo.recap: Option<String> field added to the Rust struct (mirrored on the frontend ProcessInfo interface) so the recap travels with list_bees, get_bee_status, and any new bee snapshot. set_recap() truncates to 280 chars and treats empty string as "clear"; find_bee_by_working_dir() resolves the most recently-started bee for a given cwd so the /recap POST endpoint works even when the caller only knows the working directory (not the bee_id)
beeStore.setBeeRecap preserves Map identity when the recap value is unchanged (no spurious re-renders) and trims/clears empty strings to null so consumers can use a single recap ?? null check
TerminalStatusBar slightly taller — py-1 → py-2 improves visual weight against the dense status info row
MCP project isolation enforced at the transport layer — sub-agents spawned via spawn_bee can no longer read output, write input, kill, restart, or even see status of bees in other projects. The Queen MCP server now resolves the calling bee from the TCP peer PID at SSE connect (and per-request for the streamable HTTP transport), walking the parent process chain via sysinfo until it finds a managed PID. The resolved caller_bee_id is plumbed through CallContext into every tool handler; enforce_isolation() is the single chokepoint on read_output, write_input, kill_bee, restart_bee, get_bee_status, and list_bees is filtered by the caller's project. Cross-project calls are rejected with "project isolation: bee X (project A) cannot access bee Y (project B)". External MCP clients (curl, IDE plugins) without a resolvable caller still work — no project scope is implied. New module caller_resolver.rs performs the OS-specific peer port → PID lookup (lsof on macOS, /proc/net/tcp + /proc/*/fd on Linux, netstat -ano on Windows). The pre-existing global "first running non-MCP bee" fallback that would non-deterministically attribute MCP spawns to whichever workspace happened to come first in the bee map is gone — that bug was responsible for spawning Codecs/Gemini agents into the Work workspace when the user invoked them from the Personal workspace
encoding: "screen" for read_output — full-screen TUI agents (Gemini, interactive Codex, k9s, htop) write their UI by replaying cursor moves, screen clears, and OSC color polls into the same region of the terminal. The 256KB ring buffer fills with redraw frames before the actual response can be retrieved, so utf8 reads return raw escape-sequence soup. The new screen encoding replays the byte stream through a virtual VT100 (vt100 crate, 200×50 default, cols/rows configurable up to 1000×1000) and returns just the visible terminal contents as plain text. In benchmark, a Gemini "good night" reply that was buried in 256KB of redraw bytes renders to 1935 chars of legible text
Voice transcription returns focus to the terminal — after the user clicks Stop on the voice button (or VAD auto-stops it), the transcribed text is written into the bee via write_pty and a hiveterm:focus-terminal window event is dispatched. TerminalPane listens for it and calls terminalRef.current?.focus() inside a requestAnimationFrame. The user can now press Enter immediately to submit instead of having to click back into the terminal. Skip-on-empty: whitespace-only transcriptions don't dispatch (avoiding spurious focus changes)
auto_register_codex_mcp and auto_register_gemini_mcp refresh the port on every Hive startup — the default queenMcpPort=0 means the OS assigns a fresh port on each launch. Codex's ~/.codex/config.toml and Gemini's ~/.gemini/settings.json were grabbing the port at toggle time and never updating, so MCP startup in those clients failed against a dead port until the user toggled the integration off and back on. Both functions now run alongside auto_register_mcp (Claude Code) right after the Queen binds. Decision matrix per client: section/entry present → rewrite URL with current port; absent + flag absent → first-run register + create flag; absent + flag present → respect the user's explicit disable. set_codex_mcp / set_gemini_mcp also touch the flag so a UI toggle is never overridden by auto-register on the next startup. Codex auto-registration uses pure TOML manipulation in the new mcp_register module (preserves comments / unrelated sections); Gemini uses JSON upsert that preserves all other top-level keys
Mudou
3
ProcessInfo.project_id: Option<String> — config-driven and frontend bees inherit the project from the bee_id prefix ({projectId}/{beeName}), but MCP sub-agents (id starts with queen/) lose that signal in the id. The new field is captured at spawn time so ProcessManager::project_of() can return the authoritative project for any bee. Free fn project_id_from_bee_id provides the prefix-derivation rule (returns None for queen/* and unscoped ids). Used as the source of truth for MCP isolation
build_screen_response and build_read_response extracted as pure helpers in mcp_tools.rs — both encoding paths now testable without Arc<ProcessManager>/AppHandle. Existing tool_read_output delegates to them. clamp_screen_dim enforces 1×1 minimum and 1000×1000 maximum so a runaway caller can't request a multi-MB virtual grid
MCP tool read_output description updated to advertise the screen encoding, document when to prefer it (full-screen TUIs), and explicitly note that utf8 returns the raw byte stream
Per-bee token tracking when multiple agent bees share a project — previously, two Claude (or Codex) sessions in the same cwd collided because discover_session() picked the newest .jsonl in ~/.claude/projects/<encoded-cwd>/ and assigned it to every bee, so both displayed identical token counts. The watcher now stamps each bee with started_at (millis since epoch) at spawn and pairs the oldest unclaimed JSONL with the oldest unclaimed bee within a [started_at - 5s, next_bee.started_at + 5s] bracket. Subsequent spawns disambiguate by bracket, restart re-discovers the new session by excluding the bee's previous path from the claimed set, and Gemini (which writes a single shared telemetry.log) is mapped only to the first bee so siblings don't echo the aggregate
In-app update modal with rendered changelog replaces the native confirm dialog. The modal shows the new version, the user's current version, the release notes (mini-markdown renderer covers bold, `code, text links, bullets, and headings), a download progress bar fed by the Tauri updater's Started/Progress/Finished events, and Install / Later / Skip controls. Skip persists dismissedUpdateVersion` to settings so the same version stops nagging until a newer one ships
Persistent update pill in the title bar — when an update is available and not dismissed, an amber ↓ vX.Y.Z button with a status dot appears next to the plan badge. Clicking it opens the update modal so the user can review the changelog before deciding to install
CI publishes the real changelog into latest.json — scripts/extract-changelog.sh reads the ## [X.Y.Z] section out of CHANGELOG.md, the release workflow pipes it through $GITHUB_OUTPUT into tauri-action's releaseBody, and Tauri propagates that into the updater manifest's notes field. The updater's Update.body now carries the actual release notes instead of the static "see CHANGELOG.md" link
Mudou
4
Voice record button no longer flashes on first click — the Waveform canvas had no width/height attributes, so it claimed its default 300×150 intrinsic size for one frame after mount, expanding the recording pill before useEffect resized it. Subsequent clicks were warm enough to mask the reflow but the first cold click (cpal audio engine init delay on macOS) kept the flash visible. Fixed by setting an inline style={{ width, height }} so the layout box is locked from the first paint, and switching to useLayoutEffect so the bitmap dimensions apply before the browser paints
Standardized button contrast on the amber accent — bg-hive-accent text-white (~2.4:1 contrast, fails WCAG AA) was used in five places. Replaced with the codebase-standard text-hive-amber-600 (#412402 warm-dark, ~9:1) in UpdateModal install button, VoiceButton mic/stop buttons (icon fill-white → fill-hive-amber-600 too), UpgradeModal activate, AccountTab license activate, and VoiceTab provider toggle
Token total in sidebar and status bar now excludes cache_read / cache_creation — BeeItem and TerminalStatusBar previously summed all four counters when deciding whether to render the badge and when computing the displayed total, which inflated the headline number with cached input that the user did not pay re-tokenization for. Now both surfaces use inputTokens + outputTokens; the breakdown modal still shows the cache columns separately
ProcessInfo now carries started_at: Option<u64> populated in process_manager::spawn_bee and spawn_bee_headless. The frontend already had startedAt for uptime display; the backend field unlocks server-side time-based pairing (token watcher) without round-tripping through the UI
Revert v0.16.4's Win32 Input Mode encoding — the experimental re-encoding of bare Enter/Esc as ESC [ <Vk>;<Sc>;<Uc>;<Kd>;<Cs>;<Rc> _ sequences regressed PowerShell on Windows: with PSEUDOCONSOLE_PASSTHROUGH_MODE active (Windows 11 22H2+), ConPTY forwards bytes raw to the child instead of decoding the Win32 Input Mode format, so PowerShell received the literal escape sequence as text and Enter stopped submitting commands. Restored v0.16.3 behavior: bare \r is normalized to \r\n for ConPTY (cooked-mode shells work), and \x1b passes through unchanged. The raw-mode TUI Enter/Esc issue (Claude Code, Ink agents) remains under investigation — needs hardware-level debugging on the user's Windows build to pin down whether ConPTY's input pipe is buffering, whether Ink's keypress parser is ignoring \r\n, or whether a different encoding is required for the specific Win10/Win11 build path
Windows: raw-mode TUIs (Claude Code trust prompt, Ink agents) now accept all keystrokes — the previous 0.16.2 attempt addressed only lone-CR buffering and depended on JS-side platform detection (navigator.platform, deprecated and unreliable in WebView2). Root cause was deeper: portable-pty 0.8 created the ConPTY without PSEUDOCONSOLE_WIN32_INPUT_MODE (0x4), so once a child enabled raw mode, ConPTY's input translation dropped Enter, Esc, digits, arrows, and modifier-aware keys
Switched to portable-pty-psmux 0.9.3 (drop-in fork of upstream portable-pty 0.9 from WezTerm, with PASSTHROUGH_MODE + WIN32_INPUT_MODE + RESIZE_QUIRK patches enabled). API-compatible — only Cargo.toml changed
Moved CR→CRLF normalization from TS to Rust in process_manager::write_pty(), gated by cfg!(target_os = "windows") (compile-time, immune to runtime detection failures). Implementation uses Cow<[u8]> for zero-alloc on the hot path (typing, escape sequences, paste with LF-only line endings)
Added flush() after every PTY write so bytes go through ConPTY's pipe immediately instead of waiting for the next write to nudge the buffer
Removed dead code: src/lib/normalizeInput.ts, its test file, and the IS_WINDOWS constant in usePtyBridge.ts — all superseded by the Rust-side fix
Windows: Enter key not confirming prompts in interactive CLIs — lone \r (0x0D) emitted by xterm.js was being held in the ConPTY input buffer waiting for \n, which stalled raw-mode prompts in Ink/readline-based agents (e.g. Claude Code's "trust this folder" confirm screen). Added normalizeKeyInput() utility (src/lib/normalizeInput.ts) wired into usePtyBridge.ts that translates lone \r to \r\n only on Windows. Unix PTYs unchanged (kernel handles \r→\n via ICRNL). Multi-line paste with embedded \r\n is preserved (regex uses negative lookahead).
macOS microphone permission fix — added com.apple.security.device.audio-input entitlement via new src-tauri/entitlements.plist, wired through bundle.macOS.entitlements in tauri.conf.json. Without this entitlement, the hardened-runtime signature would silently inject zeros into the audio buffer (peak=0.0000) and macOS would never register HiveTerm in System Settings → Privacy & Security → Microphone, making voice input impossible to enable on signed/notarized builds
Voice input for agent prompts — native microphone capture and cloud transcription, dictate prompts straight into any agent bee. Click the floating amber mic (bottom-right of the terminal) or press Cmd+Shift+M to record; the transcribed text is written into the bee's input without auto-submitting, so you can review before sending
New Voice settings tab with provider selector (OpenAI whisper-1 or Groq whisper-large-v3), local API key storage, "Test microphone" button with peak-level feedback, transcription language selector (11 options including auto-detect and follow-app-language), and configurable auto-stop on silence (slider from 0.5s to 5s, or fully disabled)
Animated waveform during recording — 28 centered bars driven by RMS frequency data from cpal, DPR-scaled canvas at 60fps
Voice Activity Detection (VAD) in Rust — auto-stops recording after N seconds of silence (configurable) once the user has spoken for at least 800ms, avoiding premature cutoffs and requiring zero click-to-stop
Mudou
6
Native audio capture via cpal — cross-platform microphone access (macOS/Windows/Linux) with dedicated audio thread owning the stream, lock-free VAD state via AtomicU64/AtomicBool, and WAV encoding via hound. Eliminates WebView/getUserMedia dependency
Platform-aware error messages — silence detection returns OS-specific hints (macOS: "System Settings → Privacy & Security", Windows: equivalent Windows path) via cfg!(target_os) dispatch in Rust
Whisper prompt biasing — per-language prompts (en/pt/es) with dev vocabulary ("git", "commit", "refactor", "função", "componente", …) dramatically reduce hallucinations like "ご視聴ありがとうございました" on short or quiet audio
macOS dev-mode mic permission — Info.plist embedded in the binary via __TEXT,__info_plist Mach-O section (build.rs rustc-link-arg) plus ad-hoc codesigning with the correct dev.hiveterm.desktop identifier (sign-dev.sh) so macOS TCC prompts correctly without requiring a full .app bundle
Custom range slider styling — ::-webkit-slider-thumb + ::-moz-range-thumb using hive theme CSS variables, readable in both light and dark modes
Styled language dropdown — appearance-none + Lucide chevron matching the hive theme
Windows: CMD window no longer flashes every 10 seconds — non-PTY helper commands (git branch polling, taskkill, explorer, cmd.exe env capture, reg machine GUID) now spawn with CREATE_NO_WINDOW, eliminating the transient cmd.exe popup that appeared on Windows every time the status bar refreshed git state
Mudou
1
New NoWindowExt trait (src-tauri/src/process_ext.rs) — cross-platform helper that adds CREATE_NO_WINDOW on Windows and is a no-op elsewhere, applied to all 16 non-PTY Command call sites (commands.rs, shell_env.rs, license.rs, process_manager.rs)
Image drag-and-drop now shows [Image #N] — dragging screenshots/images into agent terminals (Claude Code, etc.) now triggers proper image detection instead of showing raw file paths. Uses bracketed paste mode and writes image data to system clipboard so CLIs can recognize the image
Image paste uses bracketed paste — paste handler now sends paths via terminal.paste() instead of raw write_pty, ensuring consistent behavior with clipboard-aware CLIs on all platforms
Mudou
1
New copy_image_to_clipboard command — decodes PNG files and writes RGBA pixel data to the system clipboard via arboard (cross-platform: macOS, Windows, Linux)
Token usage tracking — real-time monitoring of input/output tokens and cost for Claude Code, Codex, and Gemini agents. Reads session JSONL files (Claude/Codex) and OTel telemetry logs (Gemini) with 5s adaptive polling
Token details modal — click token stats in the status bar to see full breakdown: input, output, cache read, cache creation, cache hit rate, and estimated cost (Claude only)
Token display settings — configurable in Settings > Sidebar: "All" (In/Out + cost), "Cost only", or "Never". Plus a master toggle in Settings > Agents
Gemini telemetry toggle — one-click enable/disable of Gemini CLI local telemetry directly from Settings > Agents (writes to ~/.gemini/settings.json)
Corrigido
3
Quick-agent not switching project — adding an agent via context menu submenu or AddCommandModal now correctly selects the target project
Agent auto-start on add — agents now auto-spawn when TerminalPane mounts instead of relying on fragile setTimeout
Token session discovery — uses file creation time (not modification time) to avoid reading wrong session when multiple Claude instances share a project directory
Mudou
3
Reduced macOS energy consumption — resource monitor uses adaptive polling (5s active, 15s idle), granular Zustand selectors in TerminalPane, cursor blink paused on hidden terminals, memoized ProjectMetrics selector
High energy consumption on macOS — resource monitor now uses adaptive polling (5s active, 15s idle) instead of fixed 3s, reducing CPU wakeups by ~40%
Excessive re-renders in TerminalPane — replaced useSettingsStore() (whole-store subscription) with granular selectors, preventing N terminal re-renders on every settings change
Cursor blink on hidden terminals — cursor animation now pauses when a terminal tab is not visible, eliminating ~1 CPU wakeup/sec per hidden terminal
ProjectMetrics re-render storm — metrics selector now returns a stable primitive instead of a new object reference, avoiding re-renders every 5s per project
Quick-agent not switching project — adding an agent via context menu submenu or AddCommandModal now correctly selects the target project
Custom title bar — native macOS title bar replaced with custom overlay showing app title, plan badge ("Free — 2/3 projects"), and Upgrade button
MCP multi-client support — Queen MCP server now supports Codex (Streamable HTTP at /mcp) and Gemini (SSE at /sse) in addition to Claude Code, with per-client toggles in Settings > Integrations
MCP auto-registration — HiveTerm automatically registers as an MCP server in Claude Code's ~/.claude.json on first launch (no manual setup needed)
Settings close button — added X button with Esc keyboard shortcut to close the Settings panel
Corrigido
3
Project reorder persistence — reordering projects via drag now correctly saves to the active workspace (was only saving to legacy openProjectPaths)
Terminal overflow — terminal content no longer overflows behind the status bar footer; added overflow-hidden to main area and xterm container
Settings modal clipping — agent edit modal now uses absolute positioning to stay within the settings panel instead of being clipped by overflow-hidden
Mudou
3
Minimum window size — enforced 680x400px minimum to prevent UI elements from being cut off
Sidebar footer cleanup — moved "Free — 2/3" badge from floating text to the title bar; simplified sidebar footer
CORS for MCP clients — opened CORS on Queen MCP server to allow local CLI clients (Codex, Gemini) to connect
Agent avatars in sidebar — agent bees now show their logo (Claude, Codex, Gemini, OpenCode, Cline) with a status badge overlay instead of the generic hex icon
Agent-specific CLI suggestions — Settings > Agents now shows relevant flags per agent type (e.g. --full-auto for Codex, --yolo for Gemini)
Dev/prod isolation — debug builds use ~/.hiveterm-dev/ for settings, license, and hooks to avoid conflicts with the installed release
Corrigido
4
Free tier limits — count only running bees (stopped bees no longer block new spawns); check limits before spawning PTY to prevent orphaned processes; enforce per-project limit on MCP commands
Project count — "Free — X/3 projects" now counts globally across all workspaces, not just the active one
Workspace switch — switching or creating workspaces now clears the active terminal panel
Spawn error UX — errors shown in the stopped overlay with an "Upgrade to Pro" button instead of hidden behind the terminal
Mudou
3
Sidebar default width — increased from 240px to 320px
Tab bar setting — moved from Sidebar settings to Appearance > Application
Projects default expanded — projects no longer collapse when switching workspaces
License lost on reboot — machine fingerprint was using MAC address which changes on macOS due to Private Wi-Fi Address randomization. Now uses stable hardware UUID (IOPlatformUUID on macOS, /etc/machine-id on Linux, Registry MachineGuid on Windows). Includes migration so existing tokens remain valid.
Mudou
2
Auto-update via Cloudflare R2 — update endpoint moved from GitHub Releases (inaccessible from private repo) to Cloudflare R2 at dl.hiveterm.com. Release workflow now automatically uploads artifacts to R2 after builds complete.
Website download URLs — now use versioned paths (dl.hiveterm.com/v0.12.0/...) for better artifact management and rollback support.
Toggle tab bar visibility — new setting to show/hide the tab bar for a cleaner workspace
macOS code signing & Linux builds — CI release pipeline now includes code signing for macOS and builds for Linux
Segurança
1
Full security audit remediation — resolved all 12 findings from security audit (1 critical, 4 high, 6 medium, 5 low): - MCP server hardened with restrictive CORS and random port binding - Path traversal fixes in config loading and image reading (canonicalize + filename validation) - TOCTOU race condition fixed — license check moved inside write lock for atomic spawn - SSE session limits (max 50) with cleanup on disconnect - Settings and port validation with atomic writes - Token file permissions restricted to 0600 on Unix - Clipboard extension allowlist and hook script injection prevention
Cline agent not appearing — new built-in agents added in updates were invisible to existing users because persisted settings overwrote the defaults without merging new entries. Added a migration in loadPersistedSettings that detects missing built-in agents and appends them while preserving user customizations.
Git Changes Panel — modal diff viewer (⌘⇧D) with GitHub-style file sidebar, syntax-colored diffs, and search (⌘F) with match navigation
Git branch in status bar — shows current branch name with dirty indicator (*), clickable to open branch on GitHub when remote exists
Open on GitHub (⌘⇧G) — detects git remote URL, converts SSH to HTTPS, opens in browser
Agent action bar — footer in Changes modal with Commit, Create PR, Revert All buttons + free-form "Ask agent..." input that sends prompts directly to the running agent
Notification click-to-focus — clicking a desktop notification (crash, agent complete, bell) focuses the HiveTerm window and switches to the agent's tab
Git hotkeys in settings — new "Git" section in Hotkeys tab showing ⌘⇧D and ⌘⇧G
Mudou
3
Git commands (status, numstat, diff) run in parallel via std::thread::scope for faster Changes panel loading
Changes button highlights amber only when working tree is dirty
Window title separator changed from em dash (—) to middle dot (·)