Release Notes

0.8.0 — 16 October 2026

Hive now speaks eight languages. This release also cleans up old sessions in one go, makes attachments easier to scan, runs a workspace's MCP servers only in the projects that need them, fits narrow windows better, and keeps Sign in in view in Agent Setup for new Copilot users.

Languages

• Hive in eight languages. Hive's window, menus and messages now come in English, Deutsch, Español, Français, Polski, Português (Brasil), 日本語 and 简体中文. Hive starts in Windows' display language when it has it, else in English; Settings → General → Language chooses another (Automatic follows Windows), and the change shows at once. Numbers, sizes, percentages and the words in dates follow the language too (1,5 MB and pon., 5 paź in Polish), and so do key names in menus (Strg and Entf in German). The installer speaks the same eight languages. --lang=<code> starts Hive in a language for one run.
• Seven of them are translated by AI, and say so. Every language but English was drafted by AI, so expect the odd awkward phrase: the list marks them (Polski (AI), Deutsch (KI)), with a note in that language under it. Each has all of Hive's text translated except two views for developers, Performance and the developer console, which stay in English.
• Help → Improve This Translation. Spotted a wrong or clumsy word? This opens the guide on GitHub for fixing a translation, or adding a language, with a pull request; each language's glossary and style notes are in docs/translating.
• The user guide stays in English for now. The documents in the Docs view (the user guide, the Agent API reference and the release notes) are in English whatever language Hive shows, though the view around them follows it; the Hive Assistant knows Hive well and answers you in your language. What Hive tells agents and the Assistant (their instructions, skills and the hive tools' replies) stays English too, so they work the same in every language. Names you gave things, and what agents wrote, stay as they are.
• The plan limits tooltip speaks one language. The status bar's tooltip mixed Hive's English with the day names of Windows' language; it is now all in the language Hive shows (Suremise/hive-desktop#1). Date format: System follows Windows' regional format, whatever language Hive shows.
• Japanese and Chinese are set their own way. Lists are joined with 、, sentences follow each other without a space, lines don't start with small kana or closing punctuation and labels break between phrases, paths show \ rather than ¥, and the "Hive can't start" window uses the language's own font.
• Clearer wording in English too. Counts in notifications, the tray menu, agents' tabs, lanes and archived cards agree with their number in every language; model and effort lists say "default" once: "Default (Opus 5.5)", "Default (Medium)", and "Claude Code's default" when Claude Code hasn't said which level that is (#892, #911, #916); and errors about cards say card in the window (the Agent API's replies are unchanged).

Long languages and narrow windows

• Text fits, or shows it's cut short. Where text was cut off or ran past its edge in a longer language, and often in English in a narrow window, Hive now shortens it with "…" (in full on hover), wraps it or makes room. The project header folds its buttons to icons, keeping the project's name and status; the agent strip fades where more agents are out of view, with a › to scroll to them; board columns, the archived cards' table, the Sessions tab, the Assistant's skills and modes, the Progress panel's runs (a reporter's name shows whole, the project name gives way first), templates and the Load template dialog fit too (#886, #890, #892, #905).
• Settings in a narrow pane. A row too narrow for its text beside its control puts the control under the text, and the section list takes less room or becomes a row across the top; the shortcuts and API prices tables, the Updates status and About fit as well (#887).
• Save stays in view in a narrow Files pane. The file toolbar's buttons show as icons (named in tooltips) when their labels don't fit, and wrap onto a second row when even those don't (#814).
• The Git tab's Changes fits a narrow window. With the Projects sidebar and the Progress panel open, the list of changes pushed the unused worktrees or the diff out of sight; it now narrows first, then moves above them (#906).
• Hive's window fits the display it opens on. Opened on a smaller display than it was last on, the window ran off the display's edge; it now opens no larger than the display and wholly on it, unless the display is smaller than Hive's smallest window (900 × 560) (#812). The developer console's window does the same (#862).

Sessions

• Clean up old sessions in one go. The Sessions tab's new Clean Up Old Sessions… button archives (the default) or deletes every session last active more than 24 hours, 7 days or 30 days ago. The dialog says how many will go before you confirm; running sessions, the one you're reading, the Hive Assistant's current conversation and any in use are always left alone, and the result says which (#895).

Attachments

• Easier to scan. Files that aren't images show their type's icon (Markdown, JSON, PDF, zip, code, text) instead of one icon for all, every tile shows its size, and a session's heading shows the days its attachments were added ("Oct 9 – today"), with a divider where each day starts (#900).

MCP servers

• Turn a server on only in the projects that need it. A server in the workspace is now available to every project and runs only where it is on: each project's MCP tab has a switch for every workspace server, and the switch in the MCP view turns one on for all projects. Adding a server from a file asks where to turn it on: the selected project, all projects, or none yet. Copy to workspace turns a project's own server on in that project. Servers you had on stay on where they were (#713).

Agents

• Agents as a table in more places. Stop All, Start New, Archive and Start New, Remove All and the quit dialog list their agents as a table (each one's state, card, worktree and what happens to it) instead of a line each, and hovering the status bar's running sessions shows the running agents the same way (#572).
• Remove All and the unused worktrees list are quicker on a busy machine. Checking which worktrees could be deleted reads the project's worktree list and main branch once, not once per worktree.

Progress panel

• Put your lanes in your own order. Drag a running lane up or down in the Progress panel's lane list (a line shows where it goes), or focus it and press Alt+Up or Alt+Down; its right-click menu has Move Up, Move Down and Usual Order. Each workspace keeps the order, after a restart too: a new lane comes in on top, and a lane that is merged or abandoned drops to the bottom, after those that stopped before it, so the latest is always last (#568).
• The filter applies to merges too. Filtered to a project (or one of its agents), the panel still showed every project's merge on top; it now shows only that project's, or that agent's (#810).

Card loops

• An agent's question for you no longer sits unseen on a card. A builder or reviewer that needs your decision now blocks the card with its question as well as notifying you, so the card shows as blocked, the Assistant's lane watch picks it up, and the agent clears it once you decide. Before, a reviewer could ask only in a comment, which nothing flagged (#840).
• A lane's merge that waits on you no longer sits unseen. A builder whose merge at the lane's end needs your OK (its CLI refused a commit, for example) now blocks the last card saying what it needs, as well as notifying you. And if a lane's cards have all passed while its builder sits idle for 10 minutes with no merge slot claimed, the Assistant's lane watch is woken: the lane's merge is stalled (#869).
• A review left without a verdict shows as stalled. A reviewer that started reviewing a card and then went back to waiting on it for 15 minutes, with no verdict and no question blocking the card, now stalls it, on the board, in the Progress panel and for the Assistant's lane watch (#843).

Hive Assistant

• A decision on a card in Review reaches its card loop. Recording your decision on a card in Review, or clearing its block, wakes neither the card's builder nor its reviewer, so the Assistant now follows it with a message to the builder, and to the reviewer when it changes what they check, instead of the loop waiting until its watch runs out.
• Apps that follow the Assistant no longer have to read its sentences. Each of its actions on the Agent API's event stream now says what it was (kind: starting a card, giving an agent a task, moving cards…) and what it was about (the project, the agent, the card or cards, the lane, the column), so a viewer such as Apiary keeps working when the sentence is reworded or translated. Nothing private is added. The API's version is now 5 (#763).

Fixes

• Sign in is always in sight in Agent Setup. A CLI installed at a long path (WinGet's package folders) widened the row and pushed Sign in out of the dialog, so a new GitHub Copilot user couldn't sign in. The path now wraps, with the full path in its tooltip (#821, Suremise/hive-desktop#2).
• What you install while Hive is open reaches new terminals and agents. Hive now reads the PATH Windows gives new programs just before it starts an agent, an Agent Setup terminal or Open External Terminal Here, so a CLI or tool installed outside Hive is found there without Check again or a restart (#821).
• The workspace no longer goes back to how it was before a change. When Hive read the projects twice close together on a busy machine, the earlier reading could finish last and show an agent list from before the change, and say again that two agents share a name. The later reading now stands.
• Retry in the Skills and Templates views shows the skill or template you then select. When their list couldn't be read, Retry in the sidebar listed them again, but selecting one still showed the failed read (Skills) or nothing (Templates) until you left the view and came back (#841).
• Copy Diagnostics no longer briefly offers an old recording. Opened again, the dialog showed the last open's developer console recording row until it had read the newest one, even when the newest couldn't be attached (nothing was attached either way). It now shows no row until the newest is read.
• Every link in the third-party notices opens a web page. Eleven packages that name their repository in a short or SSH form (owner/repo, git@github.com:…) had links that went nowhere (#788).

0.7.1 — 13 October 2026

Advanced settings

• How Claude Code's and Copilot's footers show each permission mode is in Settings → Advanced too. Hive reads an agent's mode from its footer, and watches it while switching modes with Shift+Tab: if a CLI update renames a mode there, the mode's setting fixes it without waiting for a Hive release.
• How long a stopped Copilot agent gets to exit is a setting. Settings → Advanced → GitHub Copilot → Stop: exit grace (10 seconds by default, 3 to 60): Hive ends a Copilot that hasn't exited by itself after that long, losing its token totals. Raise it if a very busy PC still loses them.

Agents

• A stopped Copilot agent keeps its token totals on a busy PC. Copilot writes a session's totals only as it exits, which took up to 2.4 seconds under load, and Hive ended it after 2.5 seconds, so on a busier machine the totals were lost. Hive now gives it 10 seconds, ending only a Copilot that doesn't exit; quitting Hive waits for it.
• Loading a template is quicker on a busy machine. Its plan and load read the project's worktrees, branches and current branch once, and each worktree's state once, instead of once per agent: about a third to a half fewer git commands, so the Load template dialog no longer takes seconds to open while tests or builds run.
• Waiting for an agent you just started on a task waits for that task. hive_wait_for_agents (and POST /v1/agents/wait) called right after starting an agent with a prompt, or giving it one in a new conversation, could answer at once that it was ready: its CLI was up but hadn't begun the task yet. It now counts as working until its CLI takes the task.

Hive Assistant

• One watch for the lanes and the cards beside them. The Assistant's watch on all running lanes can also follow cards in no lane, such as a release card: a comment, a move, a verdict or a block on one wakes it in the same line as the lanes' events, and its next lane watch keeps them. Its status says, for example, Waiting on all running lanes and #720.
• The Assistant hears when an agent waits for you. An agent showing a permission prompt or asking you something in its pane now wakes the Assistant's lane watch once, naming the agent, its card and what it asks, so it can tell you rather than let it wait unnoticed. It never answers for you.

Messages

• An agent waiting on its own timer still hears from you. An agent that ended its turn while a background command of its own runs (a sleep, a reminder to renew the merge slot) now gets messages and card events within seconds, rather than once the command ends. Agents are also told to wait for things with a card watch rather than a timer.

Terminal

• Copying from a Copilot pane works. Copilot selects text with the mouse itself and copies it when you press Ctrl+C, saying "copied to clipboard", but the clipboard kept what it had. Hive now takes a CLI's copy (an OSC 52 request) from the pane you're using, and never lets a CLI read the clipboard.

0.7.0 — 13 October 2026

Agent names

• Agent names are unique across the workspace. Two projects can no longer each have a B1, which couldn't be told apart on the board, in notifications or by the Assistant. Adding or renaming an agent with a name another project's agent has is refused, and Hive offers one with your project's initials in front (B1 in hive becomes HB1). Agents that already share a name keep working: Hive says so once, with a button to rename one.
• Name a template's agents as you load it. The Load template dialog shows each agent's name in the template beside a box for the name it will get. A name another project's agent has is marked, saying whose, and Load template stays greyed out until it's changed.
• Why a template won't load without git. A template whose agents work in their own worktrees can't load into a folder that isn't a git repository (or has no commit yet). Hive now names those agents and says what to do: make it a repository, or put them in the project folder.
• Add Agent says what hive/ in a branch name is for: it marks branches Hive made, so it can reuse and tidy them up.
• Worktree branches name their workspace and project. A new worktree agent's branch is hive/<workspace>/<project>/<agent>, such as hive/HIVE/amberlight/amb1, so several workspaces can share one repository, even with projects of the same name, without their agents' branches clashing. Branches made before (hive/b4) are still Hive's: kept worktrees on them are reused and tidied as before.

Messages

• Messages to agents are kept, and you can see where they got to. What the Assistant leaves an agent (hive_message_agent) is now a saved delivery with an id: it survives the agent stopping and Hive restarting, and its state goes from waiting to offered, received (the agent read it with its new inbox, hive_inbox) and acked (the agent says it acted on it). The pane shows ✉ 1 unread until the agent has it, and ✉ 1 overdue when it hasn't acked one within 30 minutes; the Assistant sees each one's state and times in the agent's activity. A message still waiting when the Assistant that left it stops, or Hive restarts, is refused rather than sent later, and the Assistant's activity list says so.
• Nothing Hive tells an agent is lost or counted as delivered just because it was typed. Card watch wakes, loop handoffs, your decisions, the Assistant's mode changes, a lock's go-ahead, the carry-on after a sign-in, handovers and tasks are now saved deliveries too, offered by one scheduler: each is received once the agent reads it in its inbox, and a card watch ends then, not when Hive types its wake. One Hive was typing when it stopped, or that an agent's restarted session never read, is offered once more, saying it may repeat. Tasks, handovers and your decisions are for the agent to ack; a wake or notice needs none.
• Hive types only a short line into an idle agent. Everything Hive has for an agent (a task, a card watch's wake, your decisions, a message, a notice) now arrives as one fixed line naming it by id ("[Hive] New in your inbox: 1 task (#12) (m9). Call hive_inbox to read it."), and the agent reads what it says from its inbox, labelled with who it is from. It counts as received only once the agent has read it; one let pass rings once more when the agent is idle again, and a message can be cancelled until it is read. With Settings → Agent API off, agents have no inbox: what Hive or the Assistant would give one is refused, saying so, and a card you Start yourself is typed in as your own input.
• A busy agent hears of a message between its steps, without anything typed over its work. A message the Assistant leaves a working agent now reaches it at its next tool call, as a short notice from Hive in its CLI's own hook reply, or just before its turn would end; the agent reads it from its inbox and acks it. Hive's notices say only how many items there are and their ids, so text in a card or a message that claims to come from you is read as a quote, never as your instruction, and agents ask you before deleting work or acting outside their project, whoever asks. Works in Claude Code, Codex and GitHub Copilot.
• A message about a round goes stale, not late. The Assistant can tie a message to a card's current round (hive_message_agent with card). If the card passes, leaves Doing and Review, or starts another round before the agent reads it, Hive drops the message as stale instead of delivering it late, so an agent no longer reopens a passed card for an old finding. The agent's activity and the Assistant's activity list say why.
• Agent API: POST /v1/inbox (an agent's or the Assistant's own inbox, and ack), GET /v1/projects/{name}/agents/{agent}/deliveries, and the activity's inbox replace messagesWaiting and lastMessages; the event stream's queuedMessages is now inbox. The API's contract version is now 4.

Sessions

• Transcript sizes in the Sessions tab. Every session, and every Assistant conversation, shows its transcript's size beside its context, amber past Warn when a transcript is over with a tip to hand it over to a new conversation, and — when the CLI no longer has the file.

Performance

• What each card cost. A card's panel now says what its agents used while they held it, such as cost: ≈ $1.84 · 2 rounds: its builder while it was in Doing and its reviewer from the start of each review to its verdict, as each provider reported it, with each side's part and tokens on hover. Performance has a sortable Cards table (cost, builder, reviewer, rounds) for the cards worked on in its range, a row opening its card; and hovering a lane's top line in the Progress panel shows what the lane cost, split into its cards and its overhead, what the builder and reviewer used holding no card (merges, watching, compactions), with the cost per finished card. Estimates are marked ≈ and what Hive can't know shows as Unknown, never $0. Cards worked on before this keep their cost too, worked out from their history.

Progress

• A lane's marks explain themselves. Hovering a card's segment in a lane now says what its red underline means (#650: review failed (round 4): waiting for round 5, with the reviewer's one-line gist) and why red cross-hatching is there (the card's blocked reason, or that it's On Hold, waiting for you). Screen readers get the same text.

Agents

• A Copilot agent asking whether to trust its folder shows as waiting, even on a busy PC. Copilot can draw its prompt a moment before its trust question; Hive took the prompt to mean it was ready and kept showing it ready while it asked, so nothing said it needed you. Hive now still reads the question until the agent has taken its first prompt.
• More than twelve agents in a project. Agents per project in Settings → Advanced raises the limit of twelve, up to 24, for all workspaces or just one. Above 12 it shows what more agents cost on your computer: the memory your running agents take now and how much is free. Pages, templates, the agents table and agent watches all work with 24. More than six pages show as ‹ 3 / 12 ›, with a menu of every page. Lowering it removes no agents.
• One agent's leftovers no longer take in other agents' PowerShell. With the Microsoft Store's PowerShell 7 installed, the PowerShell one agent's session started could take in every Store PowerShell started after it on the PC, another agent's or another app's: they were listed as the first agent's leftovers, and End Them ended them too. Hive now follows such a process on its own, also when it catches one in the first moments after it starts, so each agent lists and ends only what it started.
• See when a safety classifier stops an agent. When a CLI's safety classifier stops an agent's response (Claude Code's model safeguards, Codex's cyber checks), a shield beside its status says when, on which model, and the names of its last few tool calls before it (never what was stopped), with what to do about a false positive. When the CLI asks you about it, the notification and pane say why; once you've answered, the agent no longer stays Needs input but shows idle after the stop, ready for a nudge. If its model changes after one (a silent fallback), the pane shows the model it runs on now, the agent is in Agents that need you, its card gets a note, and the Hive Assistant sees it. Performance counts the stops per agent and model. Copilot's sessions don't record them, so they aren't detected there.
• A session's automatic name is the same everywhere. Until it has a name, a session is named by when it started. The agent's footer and the Sessions tab could show that time a minute apart, and a resumed session's footer showed when it was resumed. Both now show when the session began.
• What agents left running survives a restart. Processes an agent left running that Hive didn't ask you about as it quit (a quit to install an update, or with no window open), or that were listed when it crashed, are listed again the next time Hive starts, with End Them and Keep Running. Hive checks that each is still the same process, so it never offers to end a program that has since taken its place. A dev copy of Hive that an agent's test run left open now shows up there too.
• Agents stop what they start. Agents watching their own test runs with tail -f left those watchers running long after the runs ended (16 from one agent in a day). Hive's guidance now tells agents to watch a run with something that ends with it (hive-progress, their CLI's notice that a background command ended) and to end the watchers, servers and app copies they started before a card goes to Review.
• Stopping a run ends only what it started. When a command run with hive-progress is stopped, Hive ends it with what it started, as before, but now checks when each process started: a process that only shares a parent id the command happens to have reused (an older one whose parent had ended) is left running.
• An idle agent no longer stays "Working…". When Hive missed how a Claude Code agent's turn ended, the agent could show Working… at an empty prompt for good, and the Hive Assistant couldn't give it a task. Hive now takes Claude Code's own "waiting for your input" signal, a minute after a turn, and shows the agent idle again. Claude Code's compaction while idle and its recaps leave an idle agent idle.
• An agent's status follows its CLI's turns, so it can't freeze. Hive now tracks each turn by the CLI's own id. A tool result that reaches Hive after its turn has ended no longer puts a finished agent back to Working…, and a turn whose end Hive never heard of is closed from the CLI's own record (Codex's log of the finished turn) or from its idle prompt on screen (Claude Code 2.1.296 and Codex 0.162.1). When there's no sign either way, the agent shows Silent 23 min · state unknown instead of pretending to work or to be idle: nothing is typed into it, and Mark idle in its pane puts it right. Each hook Hive receives is now logged, with its ids, in hooks.log beside hive.log, so a stuck status can be traced. A card loop's stall line now says what the reviewer is really doing, not just that it's watching.

MCP servers

• Give a server to some agents only. An agent's settings now have MCP servers: All of the project's / Only these, so a server's tools reach only the agents that need it (a release-testing agent gets a sandbox server, the builders don't) and the others keep a smaller context. All of the project's stays the default; Hive's own hive server is always given; templates keep the choice. The agent's tooltip and the project's MCP tab show who gets each server, and the Agent API reports and changes it (mcpServers).

Git

• Initialise repository, in one click. A project that isn't a git repository has an Initialise repository… button on its Git tab: it makes the folder a repository on a branch called main, keeps its .hive out of git before anything is staged, adds a starter .gitignore if you choose one (minimal, Node, Python, .NET or Godot; a .gitignore of the project's own is kept), and makes the first commit, so agents can work in worktrees at once. With no name or email set in git, it asks for them and saves them to that repository only, never your global config. A folder inside another repository needs your confirmation, and one another version control system keeps is refused.
• The Git view. A project's Changes tab is now Git, with three sections: Changes (as before), History (the last 20 commits and the tags) and Branches & remotes (the branch, or a detached HEAD or a merge or rebase in progress; the remotes and their URLs; the branch it tracks and how far ahead and behind it is; the other worktrees; and who commits are made as, this repository's and your global identity, with a warning when commits would fail or would show a personal email on GitHub). From there you can switch or create a branch, fetch, pull, push, tag, add a remote or change its URL, and set this repository's identity: these are yours alone, never an agent's. Hive never fetches by itself: ahead and behind are as of your last Fetch, with its time. Switching branch or pulling waits while an agent works in the project folder or a merge holds its merge slot, and says why; agents' worktrees are shown read-only. The status bar shows the selected project's git, such as main · origin · ↑2 ↓0: click it for Branches & remotes. A project without git says so, with how to set it up.

Attachments

• Attach any file to a prompt. Drop a PDF, a log, a CSV or any other file onto an agent's or the Assistant's terminal, not only images: Hive pastes its path, copying it into the session's attachments first when it comes from outside the agent's folder (images are always kept). A chip over the terminal shows each attachment until you send the prompt, and says when the agent's CLI doesn't read that kind of file itself, so the agent opens it by its path (Claude Code reads PDFs itself; Codex and Copilot open them by path). Files over 20 MB are asked about first, and nothing attached is ever run.
• Images is now Attachments. The project tab and the Assistant view's section show every kind of file attached, with a filter by kind and previews of PDFs, Markdown, CSV and text in the viewer. Hive moves each project's .hive/images folder to .hive/attachments by itself, and old links to the images still open them. Storage and Clean Up… count and remove attachments.

Developer features

• See the text in the developer console. Settings → Advanced → Developer console: show prompt, reply and command text (off by default) shows prompts, replies, commands and tools' arguments in the console's live view, under an Unredacted banner, and ticks text for new recordings. Tokens, API keys and sign-in details stay masked whatever it says, in the text too: everything after a credential's name (--api-key, token=, password:, Authorization:) is hidden to the end of that text, and device codes are masked, agents' hive_console stays redacted, and a recording with text is still never attached to Copy Diagnostics. Copy Diagnostics masks those in its log lines now too.
• Show developer features is now the first of the Advanced settings, in Settings → Advanced's list.
• The developer console. Turn on Settings → Advanced → Show developer features, then Help → Developer Console (Ctrl+Alt+J) opens a window of its own, like a browser's developer tools: pick an agent, the Hive Assistant or Hive itself and watch each hook, status line, launch and API call arrive beside what Hive made of it, redacted. Filter by event type or provider, search, pause and clear; Hive captures a source only while the console is open on it.
• Record what a CLI sends. The developer console's Record writes a source's events to a file, redacted unless you tick prompt and reply text for that recording, with the versions of Hive and your coding agents' CLIs. Copy Diagnostics can attach the newest redacted recording to a bug report.

Advanced settings

• Fix a CLI change without waiting for Hive. Settings → Advanced now lists the details Hive relies on that a coding agent's update can break (each CLI's folder trust question, the text that means it's ready, the messages that mean its sign-in was refused, the fields Claude Code and Codex send plan limits in, Codex's title items) and timeouts for power users (each CLI's hook timeouts, the hive tools' reply timeout and Codex's timeout for them, the longest card and agent waits), plus how many lanes the Progress panel shows, which a workspace can set for itself. Search them, see which you changed with the default beside it, and Reset. They're kept in advanced.json, which you can also edit by hand: Hive reads it again when it changes. A value Hive can't use falls back to the default and says why. A value for a CLI records the CLI version it was made for and is flagged after that CLI updates, with Keep. Reset all settings resets them too. Copy Diagnostics lists what you changed (names only), and the Hive Assistant can explain and change them.
• More of what a CLI update can break is in Settings → Advanced: how Hive tells a safety classifier stopped an agent's response (Claude Code's transcript note and its question about switching models), what Claude Code's and Codex's screens show once a turn is done or while one runs, Codex's hint that it queues what you type, and the labels of Codex's /permissions menu.

Files

• A late paste no longer takes your selection. In the Files tab, a paste, rename or new file that took a moment (a big folder, a slow disk) selected what it made when it finished, even if you had clicked another row meanwhile, so Ctrl+X or Ctrl+C then acted on the wrong file. The row you clicked now keeps the selection.
• Deleting something open for a moment. Delete Project, and every other delete to the Recycle Bin, failed at once when Windows couldn't move the item because something had a file in it open, even for an instant (an antivirus scan, the search indexer). Hive now tries again for about a second and a half, so only a program that keeps it open stops it, and Hive still says so.

Agent API

• Agents can read the developer console. Settings → Agent API → Let agents read the developer console gives agents and the Hive Assistant a read-only tool, hive_console, showing what their CLI sends Hive (hooks, status line, transcript records, the terminal signs Hive reacts to) next to what Hive made of each. An agent reads only its own sessions; the Assistant reads its workspace's agents and Hive's own events (API calls, merge slots, watches, Progress). Hive captures only while the developer console is open, and what it captures is redacted: no prompt or reply text, commands, tokens or sign-in details. Off unless you turn it on.
• Viewer tokens, for apps that show Hive. An app that only shows what your agents are doing, such as Apiary, can now get a read-only viewer token: Settings → Agent API → Viewer tokens makes one, named for its app, that reads everything in your workspaces and follows its events but can't change anything (every change is refused); Revoke cuts that app off at once. Such an app finds a running Hive by itself from %APPDATA%\Hive\agent-api-discovery.json, which holds Hive's API address (never a token) while the Agent API is on.
• More events, and the Progress panel's runs. The event stream (/v1/events) now also sends progress runs, merge slots, Hive's notices, the Hive Assistant's actions and the agents it adds, so an app needn't poll for them. GET /v1/progress lists the Progress panel's runs. Nothing private goes out on the stream or in the runs: no prompts, messages, questions or replies, no CLI output, no command lines (their arguments can hold a token) and no notice text; a run's title goes out only when it was given on purpose (hive-progress --title), and a run titled by its command line is just "a command". An agent's state on the stream now says that a message waits for it or it asked something, not what.

0.6.0 — 11 October 2026

MCP servers

• Add MCP Server from File. The MCP view adds a server from a .json definition or a .zip of a server's folder (its code and a hive-mcp.json definition), as the Skills view adds a skill from a file. Hive shows the name and what the server runs (its command and arguments or address, and its environment variables by name) before adding it, asks before replacing a server of the same name, and adds it turned off. Enabling any server now shows what it runs and asks first, and a server's view shows it too. Zips are unpacked safely: nothing in them runs, and paths outside the server's folder, links and oversized archives are refused.

Agents

• A kept worktree is used again. Adding an agent with a new worktree named like an unused one (a new B4, with hive/b4 kept from the last) now gives it that folder instead of making hive/b4-2, when it is merged and clean: Hive brings it up to main first, and runs the setup command again only if a lockfile changed. One that holds work is left alone, and Add Agent offers to reuse it with its work. The Agent API (hive_add_agent) does the same, can take an unused worktree by existingWorktree, and says which happened.
• A project's agents as a table. Hovering a project's tile in the compact sidebar, its agent count or status in the full one, or the status in the project's header shows its agents as a table: each one's state, its card (cut to one line) and what it waits for, a lane's builder beside its reviewer, instead of a wall of wrapped lines.
• Hive offers to end what a stopped agent left running. A preview server, a tail -f or a shell an agent started in the background could keep running long after the agent stopped, holding the project's folder ("in use", so it couldn't be renamed) or a port. Hive now keeps track of every process an agent's session starts, and when the agent stops with some still running, a notification lists them with End Them and Keep Running, and the agent's pane header shows ⚠ n left running until you decide or they end. Settings → Agents → Processes left running can end them every time instead, and quitting Hive asks about any still listed. Hive only ever ends processes the agent started. That includes what an agent runs through the Microsoft Store's PowerShell 7 (Codex uses it where it is installed), whose processes Windows starts outside the session's tracking.

Progress panel

• Lanes: four by default, drag for more. The lanes at the bottom of the Progress panel now show up to four (six took too much room from the runs) and scroll after that. Drag the line above Lanes up to show more, a whole lane at a time, or down for fewer; double-click it for four again, or use Up and Down on it from the keyboard. Each workspace keeps the height, folded or not and after a restart.
• Honest run states. A run queued for a test slot shows what it waits for and how long it has waited, never a time left ("about 0 s left" read as nearly done), with a still, hatched bar. A run past its estimate says "finishing", then "longer than expected". A command stopped with Ctrl+C shows as Stopped, and one whose hive-progress was ended with it (an agent stopping a preview server) as Ended without a report once its process is gone, instead of "stopped reporting".
• Merges and lanes at a glance. While an agent merges, a teal card at the top of the Progress panel says who is merging which cards, the branch going in (hive/b1 → main), the step its checks are on, how long so far and left, and who is waiting; it folds to one line. When builders and reviewers work through cards in lanes, the bottom of the panel shows each lane as a row: the pair, the card it is on, how many are finished, and a bar with a segment per card, coloured and patterned by where the card is (to do, doing, in review, passed, done, blocked or waiting for you, a failed review). Hover a segment for the card and its round; click it to open the card. Long lanes fold their finished cards into one "✓ n" segment, and the lanes fold down to their header. With no merge and no lanes, the panel is as before.

Providers

• Price tables follow the CLIs. Each provider's API prices now lists the models its CLI has, so a model a CLI update adds shows without a Hive update, marked No price: enter one until you do; a priced model the CLI no longer lists stays, greyed, for older sessions. Claude Code (and Copilot) have a new Session costs choice: the CLI's own reported cost (the default, as before) or My prices, your table for every session. Codex's table says it is the only source of its estimates, with a link to OpenAI's pricing.
• GitHub Copilot's Allow all needs no opt-in; Autopilot still does. Allow all (approves every tool, path and web request without asking, in the agent's folder and outside it) is now offered like any other mode, with no red warning, so Copilot has a mode for working with few interruptions. Autopilot stays behind the opt-in, now Settings → GitHub Copilot → Enable the Autopilot option; if you had ticked it, Autopilot stays available. Checked with Copilot 1.0.93: once a turn has run in Autopilot, Copilot approves everything for the rest of the session, even after you switch out of it, and Hive's mode badge then shows Allow all. Autopilot's description and the user guide now say so.

Lanes and loops

• Card reports lead with a one-line summary. Round reports and review verdicts start with one line ("Round 2 ready: …", "Round 1: FAILED: 2 findings: …") and the detail follows, and agents and the Assistant can read just those lines of a card's comments, so following a loop no longer means reading every long report in full.
• The Assistant watches whole lanes. Instead of a list of at most 20 cards it swapped in and out (and once missed a card waiting for review), the Assistant waits on every running lane at once, lanes started later included, and is woken only for what needs it: a failed review, a blocked or stalled card, a lane ready to merge, a lane ended. Routine moves no longer wake it.
• The Assistant can leave a message for a busy agent. Agents in a card loop are nearly always waiting or working, so the Assistant couldn't pass on your answers or re-brief them, and you pasted its messages into their panes yourself. Now it leaves the agent a message: one waiting on its cards gets it at once and keeps waiting afterwards, and a busy one gets it as soon as its turn ends. The agent's pane shows ✉ 1 message waiting until then; hover to read it, click to cancel it.
• Your card loop defaults are settings. Settings → Workspace → Card loop rounds and Card loop wait set how many rounds a card gets and how long agents wait before telling you, for every loop in the workspace that doesn't say otherwise; a project can set its own in its Settings → Card loops. They start at 10 rounds and 6 hours (they were 5 and 2 hours), and what you say when you start a loop still wins.
• Hive wakes reviewers and builders itself. When a card arrives in Review, Hive wakes its reviewer; when the review gives its verdict, Hive wakes the card's builder, even when that agent didn't set a wait, so a loop no longer stalls because one agent forgot. It only does this for cards with a reviewer (the reviewer of the card's lane, or the agent that last reviewed it), only while the agent is idle, and never when the agent's own wait already tells it. The card's history says whom Hive woke. Turn it off for a project in its Settings → Card loops.
• Hive knows the card loops it runs. When two agents work through cards as builder and reviewer, the loop is kept as a lane: its builder and reviewer, its cards in order, its rounds and its wait. The builder registers it as it starts and the reviewer joins it, with nothing for you to set up. Hive follows the board from there: the lane is on the first card not yet passed, paused while that card is blocked, and ended once its cards are all Done. A lane whose builder or reviewer you remove is abandoned. The Assistant can list the lanes and start one when it briefs a pair, and scripts read them through the Agent API (/v1/lanes). A lane is only a record: an agent working on a card on its own has none, and a loop works the same without one.
• Agents write tests that hold up on any machine. The bundled work-on-card and review-agent-work skills now say that tests wait for what they expect to see, never for a length of time: no sleeping and then checking, no single read just after an action, no raised timeout as a fix, no test depending on another's leftovers. Reviewers fail a new or changed test that does, since it fails on a loaded or slower machine.
• Reviewers don't repeat the builder's test runs. In a card loop the builder's run record (which checks ran, on exactly which code) is what the reviewer checks, rather than running everything again: the bundled review-agent-work skill now has it run the quick checks, the tests the change affects and its own probes of the riskiest parts, and a whole suite again only when the record is missing, for other code or doubtful, or the change is to shared test code, saying why. Reviews take less time and fewer tokens, and don't load the machine while other agents test.
• The Assistant runs a team out of the box. In Coordinator mode, a new skill that comes with Hive, run-lanes, has the Assistant plan lanes by the files each touches, brief every builder and reviewer in full, and watch each lane to its end, checking the board whenever a lane changes and what agents report, instead of leaving lanes to "tell you". It does this when a project has a team (three or more agents, builder and reviewer pairs, or agents in worktrees); with one or two agents it stays light and at most suggests more structure in a line, and you can still use card loops and everything else. If it starts a lane without watching it, Hive's reply reminds it, and when a card wakes it while lanes are running, Hive reminds it to watch them all again.

Fixes

• Sizes and token counts read cleanly. The Performance view showed an average under 1 KB unrounded ("160.17241379310346 B") and billions of tokens in millions ("12904M"). Sizes are now whole bytes ("160 B") and token counts have a billions unit ("12.9B"), everywhere Hive shows them; a value that would read "1000k" or "1024 KB" shows in the next unit ("1.0M", "1.0 MB").
• Wakes are sent when Claude Code is busy for a moment. About one wake line in fifty sat in an agent's or the Assistant's prompt unsent, even after Hive pressed Enter again, until you pressed Enter yourself. Claude Code, busy for a moment, read Hive's line together with the keys around it as a paste and kept those keys as text. Hive now puts its lines into Claude Code as a paste, which Claude Code reads as text however it arrives. A long line shows as [Pasted text #1] in the prompt until it is sent.
• Board changes stay quick on a big board. Each new card, move or batch change read every card file, one at a time, while other changes waited: on a board of 500 cards, a new card took about 0.2 s and a batch of 30 changes over 5 s. Placing a card now reads only the cards changed since, and the board is read several files at a time: about 0.1 s and 2 s.
• A reviewer waiting for a card that came back to Review is woken. A reviewer told a card was in Review, which then went back to Doing (blocked to ask you, say), and asked again to wait for it to be in Review, was never woken when it came back: on 10 Oct three reviewers sat waiting for five hours on cards already in Review. Hive now wakes them. A card that waits in Review for 15 minutes while its reviewer waits for it, without its review starting, now shows as stalled on the board and in its lane's row in the Progress panel, and an agent whose turn ended while its background tasks run (a long test run) is no longer reported as not watching its card.
• Sessions are found when a folder has another name. A project or worktree reached through a junction or symlink, or by an 8.3 short name (C:\Users\RUNNER~1\…), could lose its sessions: Claude Code files its transcripts under the folder's real path when started through a junction, and Codex and Copilot keep the name they were started with, so a session started by one name and looked for by another had no transcript, usage or entry in the Sessions list. Hive now finds a folder's sessions by where it is, whatever its name, and Repair, after a move, copies Claude Code's transcripts and memory from every name they were kept under.
• Escape closes a new, empty card. Pressing Escape just after opening the first New Card of a session could ask "Discard unsaved changes?" about a card with nothing in it. It closes the dialog now.
• The Load template dialog shows the agents as a table. Each new agent is a row with its coding agent, model, effort, mode and where it works, in the template's order. Settings left to the defaults say so in a muted (default), and a new worktree shows its branch, with its folder on hover. Before, each agent was a long line that wrapped mid-path, which was hard to read for a template of 12 agents.
• Save in a template's agent saves the template. Editing a template, then pressing Save in one of its agents' settings, changed nothing until you also pressed Save Template, so the change was easy to lose. The agent's Save now saves the template, and the editor stays open. Other changes not saved yet are marked in the editor and in the list, and are kept if you go to another template and come back. Quitting or reloading asks about them, as it does for unsaved files.
• A renamed project keeps its cards. Renaming a project folder left its cards under the old name. They vanished from the project's board and Archived, and showed their agent as "(removed)". Repair moved the project but not its cards. Hive now names a project by an id kept in its own .hive folder, so cards, lanes and agents' watches follow a rename or a move, even one made while Hive runs, and watches also follow a move of the whole workspace. A merge left half done when Hive closed is reported in the project as it is now. A renamed project also keeps its handovers (so an agent picking up finds the latest one), its Performance history and kept comparisons, its Progress runs and its tabs' view settings. They also stay with their own project when two projects swap names or a new folder takes the old name. Cards saved before this version still name the old folder: for a project renamed since, the banner offers Repair…, which gives them to the project they belong to now. That is the project renamed from that name, or else the one whose agents worked on them.
• A project renamed twice in quick succession keeps its settings. If Hive failed once to carry a project's per-project settings and Progress runs over to a new folder name (a file in use, say), and you renamed the folder again before it tried again, the retry could act on the older name and record a rename back to it. The settings and runs then followed a folder that no longer existed. The retry now uses the folder names as they are now.
• The board always hears that a project was renamed. If Hive happened to read the projects part way through a rename, or a project folder was moved out of the workspace and back under another name, the open board kept showing its cards under the old name until something else made it read them again. It now updates.
• A card started on an agent with extra arguments no longer sits idle. When an agent's extra arguments ended with an option that takes a list, such as Claude Code's --allowedTools mcp__hive Bash or Codex's --image a.png, the CLI read the task Hive starts it on as one more value of that option: the agent opened at an empty prompt and the card stayed in Doing. Hive now passes the task after --, which ends the options. Extra arguments can no longer hold a bare -- of their own: saving one is refused, saying why, and one saved before stops the agent from starting until you remove it. An option whose value is -- is written --option=--.
• A GitHub Copilot agent no longer shows Ready while it works on a new conversation's first prompt. When you refused one of Copilot's prompts and then started another conversation (/new, /clear, /resume) and prompted straight away, the old conversation's end could be read after the new prompt, so the agent showed Ready while it worked. The old conversation's end now counts as older than the prompt that left it.
• The Assistant at Look and advise can list the card loops. Asking it "which lanes are running?" needed Control agents, though it only reads them; it now answers at every level (starting or ending a lane still needs Control agents).
• A blocked card in a loop is reviewed when it comes back. When a builder blocked its card to ask you a question, its reviewer dropped the card and moved on to the next one, so once you'd answered and the card was back in Review, nobody reviewed it and the loop stalled. A blocked card now stays in its loop, paused: the reviewer keeps waiting for it and reviews it when it returns.
• Hive warns of a stray .codex folder. A .codex folder an agent left scratch files in could stop every Codex agent's commands in that project ("setup refresh had errors"). Hive now warns when a project or worktree has a .codex folder holding something other than Codex's configuration (its config, custom agents, rules), and agents are told to keep scratch, probes and logs in a temp folder, out of projects and their .codex, .claude and .agents folders.
• A restart no longer switches the Assistant to Codex. Just after Hive started, the default provider (Automatic) could pick Codex for a moment when Codex answered before Claude Code had been found, so resuming the Assistant's Claude Code conversation was refused ("the Assistant now runs Codex") until Hive had finished looking. Automatic now waits until Hive knows which CLIs are installed (Settings shows Automatic (checking…) meanwhile), and a conversation always resumes in the provider it ran in, in the Assistant or in an agent whose provider you've changed since: a provider setting is for new conversations.
• The Progress panel's lanes show what is happening now. A merged or abandoned lane is greyed and still, and shows its cards as they were when it stopped, so a card another lane has taken on no longer looks worked in two lanes at once. Running lanes are listed first, six fit before the list scrolls, and a lane shows the card its builder is actually working on: one waiting for your answer no longer reads as paused once the builder has moved on to the next card.
• One fewer third-party program in the installed app. Hive no longer ships resources/elevate.exe, a helper the installer builder adds for updates that need administrator rights. Hive installs per user and its updates never need them. Its notice is gone from the third-party notices.
• Open file in the Changes tab works for every worktree. For an unused worktree outside the workspace (one made with git worktree add elsewhere), Open file above its diff did nothing but report an error. It now opens that worktree's copy of the file, as it does for agents' worktrees and the project folder.
• The Assistant moves a whole set of cards in one go. Moving or labelling many cards in one step (up to 100) counted each card toward the Assistant's 30 changes for a message, so moving 37 cards to the next column stopped part-way and asked to go on, which was easy to miss. The step now counts as one change, however many cards it has; each card is still listed on its own in Done by the Assistant.
• The same folder by another name is the same folder. A workspace reached through a junction or a short (8.3) name, such as C:\Users\RUNNER~1\…, could be opened in a second window while it was open in the first, and a workspace reopened by another name looked moved, offering Repair…. Removing a project could pass a worktree that is another project of the workspace when its path was written another way, and resuming a session was refused ("This session ran in …") when its worktree's path had changed form. An Agent API call naming its workspace by another name in the X-Hive-Workspace header wasn't found, and a locked file could be named as ..\..\RUNNER~1\… rather than by its path in the agent's folder. Hive now compares these by where the folders really are.
• Your answer reaches an agent that asked before starting its card. An agent with a question about a card still in Todo could ask you and leave the card there with nobody on it, so the decision you recorded for it was nobody's and Hive didn't tell the agent. Agents now take the card (to Doing, themselves on it) before they ask, so Hive tells them of your decision as it does for a card they're working on.
• Copy Diagnostics leaves out where Claude Code and Codex were looked for. When a Claude Code or Codex Hive tried didn't report its version, the log line kept the folder it was in and what it printed instead, so a folder with a space in its name, or a private name in the message, could reach a bug report. Both are now left out, as GitHub Copilot's already were; the line still says a copy didn't report its version.

0.5.1 — 9 October 2026

Fixes

• Copilot panes are drawn at the right width from the start. A GitHub Copilot agent's screen could load garbled (its input box and "Loading" lines drawn at about two thirds of the pane's width, several frames stacked, stray characters at the old edge) until you resized the window, and a mode switch could do the same. Each agent's terminal now starts at its process's size, a restarted agent starts at its pane's size, and Hive has Copilot redraw at its size once it has started and once it is ready, as a resize by hand did. Claude Code and Codex agents get the first two as well.
• Copilot's "Manual Approval" in Accept edits is explained. Copilot's own footer still says Manual Approval in Accept edits, since Hive approves the edits for it; the mode badge's tooltip and the user guide now say so, so it doesn't look as if the mode didn't take.

0.5.0 — 9 October 2026

Hive Desktop

• Hive is now called Hive Desktop. The window title, About, the status bar, notifications, the tray, the installer, the Start menu and desktop shortcuts and the entry in Installed apps say Hive Desktop (in sentences, Hive is still Hive). Nothing you have moves: the update renames the shortcuts, and Hive stays installed where it was, as Hive.exe, with your settings and conversations in %APPDATA%\Hive; taskbar pins and notification settings carry over.

Providers

• New agents use the coding agent you have installed. Default provider (Settings → Providers) now starts as Automatic: new agents use the first provider that is turned on and installed (Claude Code, then Codex, then GitHub Copilot). If you only use Codex or Copilot, an empty project no longer says Claude Code is missing, and New Session adds an agent with the CLI you have in one click. Hive warns only when no coding agent is installed, or when a provider you chose or an agent uses isn't: a provider you turned on but don't use no longer brings a banner or opens Agent Setup. If the default you chose is turned off, Add Agent says so instead of quietly picking another. If your default was Claude Code, it becomes Automatic, which still picks Claude Code wherever it is turned on and installed; a Codex or Copilot default stays. Agents you already have keep their provider.
• GitHub Copilot agents. Agents can now run on the GitHub Copilot CLI (copilot), beside Claude Code and Codex: turn it on in Settings → GitHub Copilot, then choose it when you add an agent. They get what the others get: statuses from Copilot's hooks (working, asking for permission, asking a question, finished), file locks (an edit of a file another agent is changing is refused, or handed to Copilot's own approval dialog in Ask me), the hive tools and Hive's guidance, the workspace's skills, sessions Hive names and resumes, and the mode in its footer (Accept edits, the default, Ask, Plan, and Autopilot and Allow all when turned on; Shift+Tab switches to Plan and Autopilot and back while it runs). Tokens other tools leave in your environment (GH_TOKEN, GITHUB_TOKEN) are kept from it, so it uses its own login, or COPILOT_GITHUB_TOKEN if you set one for it; the first start in a folder asks whether you trust it, in its terminal. Copilot Free runs the Auto model only; the account's models are read from the CLI. Stop ends a Copilot agent with Ctrl+C, as you would, so Copilot keeps the session's token totals: it takes a second or two, and Hive ends it after a few seconds if it doesn't exit.
• Hive's own provider badges. Claude Code, Codex and GitHub Copilot are now shown with badges of Hive's own instead of their makers' logos: a coral circle with a tilde for Claude Code, a teal rounded square with braces for Codex, and a violet diamond with a double chevron for GitHub Copilot, always beside the provider's name (or as its tooltip). On the status bar they're drawn in its colour, and their outlines tell them apart. About, the README and the user guide say Hive Desktop is independently developed and isn't affiliated with, endorsed or sponsored by Anthropic, OpenAI or GitHub.
• Agent Setup for GitHub Copilot. Its tab finds the Copilot CLI (installed with WinGet or npm), shows its version, and installs or updates it with the same installer. It says who Copilot is signed in as and how: a COPILOT_GITHUB_TOKEN you set, its own copilot login, or the GitHub CLI's login, which Copilot falls back to. Signed out, Sign in opens copilot login in a terminal for you to sign in. Hive reads only account names, never a token. A note says Copilot Free uses Auto only; choosing a model needs a paid plan. The user guide's new GitHub Copilot section covers installing, signing in, AI credits, and what Copilot Pro's "third-party agents" are (cloud agents, not Hive's Claude Code or Codex).

Worktrees

• The Changes tab leads with your unused worktrees. Its picker now starts with Unused worktrees, then the worktrees agents work in, then the project folder, last. A project with worktrees opens on Unused worktrees; one without opens on its folder. Each unused worktree says who made it: was B4 (the agent that last worked there), made by Hive, or not made by Hive (one an agent made itself with git worktree add, say). Listing them takes a few seconds, so the page shows a spinner and says what Hive is doing the first time, then shows the last list at once while it checks again, with Refresh to look again. Each page says in a sentence what it shows. Agents are now told to make no git worktree of their own to work in unless you ask (Hive's locks, Changes and merging follow the folder Hive gave them), and to tidy a subagent's worktree before they finish.

Fixes

• A CLI you install from Agent Setup is found straight away. Installers add their folder to your PATH in Windows, which programs already running don't see, so Hive didn't find GitHub Copilot (or another CLI) it had just installed until you restarted Hive. Now, whenever Hive looks for the CLIs (when it starts, when an install or update ends, and with Check again), it also reads the PATH Windows gives new programs and adds its new folders to its own. Agent Setup shows the CLI installed, and agents you start from then on can use it, without restarting Hive (the installer's advice to restart your terminal is for terminals already open). The same goes for Git: install it, then Check again.
• Copilot's danger setting names both modes it unlocks. Settings → GitHub Copilot called it "Enable the Autopilot option", though it also unlocks Allow all. It now reads Enable the Autopilot and Allow all options, with its description and confirmation to match, and turning it off moves projects and agents using either mode back to Inherit (Allow all ones used to keep it). Claude Code's and Codex's read as before.
• Refusing Copilot's path prompt no longer leaves the agent waiting. When a GitHub Copilot agent asked for a path outside its folder ("Path permission needed") and you pressed Esc, Copilot ended the turn without telling Hive, so the agent showed waiting until your next prompt. Hive now reads the turn's end from Copilot's session and shows the agent ready.
• A Codex agent you signed in again for stays working. When Codex refused a turn for an expired sign-in and you sent the next prompt just as Hive was reading that refusal from Codex's session, the late read could still mark the working agent Needs sign-in. A refusal read before a newer prompt or other hook no longer overrides it.
• Codex's sandbox setup and permission switches no longer stall on a busy computer. Hive types /permissions into Codex for Set up in Agent Setup and when you switch a running Codex agent's preset. If Codex was busy just then, the command could stay in its input, never run: the setup prompt never came up, or the switch was undone after a while. Hive now also waits while Codex's screen shows it is busy holding what was typed ("tab to queue message"), and if Codex still holds the command once it is free, presses Enter again (twice more at most). A command Codex has taken is never sent twice, and if Codex stays busy, asks you something or stops meanwhile, Hive types nothing more and says the switch wasn't made.
• Agent Setup's chips line up. The Signed in chip's account icon was drawn at full size, so it stood taller and sat higher than the chips beside it on the Claude Code and Codex tabs. The chips now share one height and centre line, and wrap neatly. The Git for Windows and Install buttons' icons are back to button size too.

0.4.1 — 8 October 2026

Task board

• Your answers reach an agent waiting for you. When an agent asks you something and its turn ends, and you answer through the Assistant (which records your answer as decisions on the card), the agent used to hear nothing until you pasted the answer into its pane. Now Hive types one line into it once nothing is in the way ("#392 has 2 new decisions from the user: read them and carry on"): decisions recorded together come in one line, never while it works or you type in its terminal, and the card's history says whom Hive told.
• Save keeps a decision you're writing. In a card, Save used to post a comment you were writing but drop a decision you were writing. It now records both (and a change you were making to a decision). Add Decision and Comment are highlighted while their box has text, and say that Save adds them too.

Models

• Claude Haiku 5.5 (released 7 October 2026) is priced: $0.10 per million input tokens and $0.50 per million output (prompts up to 100,000 tokens; longer prompts cost more and aren't estimated yet). Its cache prices follow Anthropic's pricing page. The haiku alias now means Haiku 5.5, as Claude Code's does, and Haiku 5.5 is in the pinned models; Haiku 4.5 stays in the older versions, and its sessions keep their price.

Skills

• A card loop moves on when its card is taken away. If a card an agent is waiting on in a card loop is given to another agent, reviewed by another, or blocked (for your decision, say), the agent is told once and carries on with its next card, instead of waiting until you cancel its wait by hand. A wait on several cards keeps the others, and the card's history says whose wait ended.
• A card loop never waits for what has already happened. A builder that started waiting for its card to reach Passed after the card was already there waited until its wait ran out (six hours, once), holding up the merge. A wait for a card to reach a column it's already in now ends at once. A reviewer waiting for a failed card to come back asks for the card's next move into Review (fresh in the Agent API), so it still waits while the card sits in Review.

Hive Assistant

• A tidier Assistant panel. The header no longer repeats the status with a dot: the Status line under it says it, after the Assistant's own icon (a hive cell with a crown), in the status's colour with an agent's status dot on it. The same icon marks the Assistant on the panel's folded strip, its runs in the Progress panel, the inbox and the Workspace Overview. The cards it waits for are the same coloured card chips as everywhere else (click one to open the card). Done by the Assistant now sits between the status and the workspace, with the same last 3, Show all and fold. Each project in the workspace list folds to one line (its name, the dot of the agent that most needs you, and how many agents run), and stays folded for that workspace. Each agent in the list has a faint backing in its status's colour, a soft wash while it works and a light hatch while it waits on cards or for you, so you can see at a glance which ones are busy.
• Several cards in one change. The Assistant can move or update up to 100 cards in one step: the same column, position, blocked, labels or agent on each. Each card is checked and changed on its own, with its own history line, and one it can't change (in Doing with another agent, say) is left as it was while the rest change. Its 30 changes per message count each card; a batch that doesn't fit changes none of them, and says how many would fit. Agents can batch the cards of their own project too.
• The Assistant waits on agents without checking in. Asked to tell you when an agent is done with work that has no card (a build, a test run, a push and its CI), the Assistant used to check on it every few minutes, each check costing context and tokens. Now it starts an agent watch and stops: Hive wakes it with one line when the agent finishes, needs you, or stops ("B6 (hive) finished: "The installer is built…""), and its status line says Waiting for B6 to finish meanwhile. Agents can watch another agent the same way (hive_wait_for_agents with wake).
• The Assistant knows you by name. When Hive tells the Assistant you replied to an agent in its pane, the line now says "User replied to B1…", or uses the name you set in Settings → General → Your name ("Darren replied to B1…"). The name goes into the Assistant's conversation only if you set one.
• The Assistant hears when you answer an agent yourself. When you reply to an agent in its pane while the Assistant is watching it (or its card), the Assistant is told the first line of what you sent, so it doesn't ask you again and can record your decision on the card. Settings → Assistant → Tell the Assistant when I reply to an agent turns it off.
• Each new lane in a new conversation. The Assistant can give an idle agent its next task in a new conversation in one step: its current conversation ends (it stays in the Sessions tab) and a new one starts on the task, with no handover. It does this when it starts an agent's next set of cards in a card loop, and keeps the same conversation within one. Shorter conversations cost much less, since every request re-reads the whole conversation. It is refused, as other tasks are, while the agent is working, starting, waiting on its tests, waiting for you or watching a card, or just after you typed in it.

Hive

• A clear (×) button in every search and filter box. The board's search, Settings, the Files and Sessions tabs, the sidebar's filters, tables' filters, Tips, Keyboard Shortcuts and the command palette show an × while they have text: click it, or press Escape, to clear the box. In a dialog or the palette, the first Escape clears the box and the next closes it.
• Easier-to-read secondary text. Timestamps, counts, card numbers, hints and other grey text, and amber text and icons, now meet the WCAG AA contrast level (4.5:1) on every background in both themes, hovered and selected rows included. Status dots meet 3:1, and a pulsing dot now changes shade instead of fading. Most of the change is in the light theme, where faint text was hard to read.

Worktrees

• Unused worktrees live in the Changes tab. Worktrees no agent uses now (kept when you removed their agent, or from another template) are listed in the Changes tab's folder picker, under Unused worktrees, next to the agents' worktrees. All unused worktrees has the same Remove, Remove all merged, Remove anyway… and Give to an agent… as before, and Open shows one's changes like an agent's worktree, with Merge to bring its work onto main (and remove it afterwards, if you like). The project Overview keeps a one-line count with a link.

Terminal

• Card numbers are links. When an agent or the Assistant mentions a card in its terminal, like "#383 is back On Hold", hover the number to see the card's title and column, and Ctrl+click it to open the card, as you would a file path. Only numbers of cards on your board are linked.

Fixes

• The command palette's highlighted command shows its shortcut. In the light theme the shortcut of the highlighted command was white on white, and in the dark theme hard to read; it now reads clearly on the highlight in both.
• A wake that isn't taken no longer sits unnoticed. Sometimes a CLI (seen with Claude Code, the Assistant's too) isn't reading its input when Hive types a wake into it, and the line sits in its prompt. Hive already pressed Enter once more; now, if it still isn't taken, the agent's pane (or the Assistant's status line) says A line Hive typed wasn't taken and you're told once, so you can click its terminal and press Enter. Hive types nothing more into it until then, so a later task isn't lost behind it, and the log says when each wake was taken.
• Sonnet 5.5's cache reads are priced at $0.10 per million tokens, Anthropic's published price (5% of input), not $0.20. Estimated costs of Sonnet 5.5 sessions that Hive prices itself come out lower where most of the input is cached. A price you set for Sonnet 5.5 in Settings → Claude Code → API prices is kept.
• Hive opens when its install folder's permissions don't cover sandboxed apps. If the folder Hive is installed in had an entry for a Windows app package but none for ALL APPLICATION PACKAGES, Windows denied Hive's sandboxed windows access to it and Hive closed at start with nothing on screen. The installer and updates now give ALL APPLICATION PACKAGES read access to the folder, and Hive checks at start: it adds the entry to a folder you can change, and for one you can't, it says so with the command to run and a Copy Command button instead of failing silently.

0.4.0 — 7 October 2026

Hive Assistant

• The Assistant's status, at a glance. A line under the panel's header says what it is doing: Status: Working, Idle, Waiting for you (highlighted), or the cards it is watching and until when, each card a link that opens it.
• "Done by the Assistant" moved under the terminal. It shows the last 3 changes at a fixed height; Show all scrolls through every one in the same space instead of pushing the terminal around, and clicking its title folds it away (remembered for the workspace).
• New Mode… no longer opens again after Enter. Pressing Enter to create a mode (or in any dialog opened from a button that you confirm with Enter) left a new, empty dialog open, because the same key press reached the button that opened it.
• A resumed Assistant uses its current mode. Resuming the Assistant's conversation kept the instructions it started with: Claude Code sends a conversation's first system prompt again on every resume, so a mode chosen while the Assistant was stopped, Restart in This Mode…, or a Hive update's new instructions didn't reach it. Hive now resumes it with its current mode's instructions, and when you switched mode while it wasn't running, tells it the new mode in its conversation once it is ready, as a switch while it runs does.
• See the images you've pasted to the Assistant. The Assistant view (the robot in the activity bar) now has Modes, Conversations and Images. All Images shows every screenshot you've pasted or dropped into its conversations, grouped by conversation, with the same viewer as a project's Images tab; click a conversation's name to read it, or move a whole conversation's images to the Recycle Bin at once. Its conversations use the new Sessions tree. Also in the command palette: Show Assistant Images.
• The Images views say what they hold. A line under the heading of a project's Images tab and the Assistant's All Images says whose images they are: those pasted or dropped into the project's agents' sessions, or into the Assistant.
• The Assistant's workspace overview lists each project's agents under its name, with a small indent, however wide the panel is. They used to start a third of the way across, drifting right as the panel widened.
• The Assistant's personas are now working modes. Coordinator (the default) runs the work across your agents, Planner shapes work into questions, lanes and cards, QA triager turns a bug report into a card with evidence, and Release manager takes a release through its checklist. They replace the character personas: an Assistant that used Overseer, Orchestrator or Reviewer moves to the closest mode, and personas you wrote or edited stay. Switching mode from the panel's header now takes effect at once and keeps the conversation (Hive tells the Assistant the new mode), instead of starting a new one; Restart in This Mode… resumes the conversation with the mode's full instructions. The Assistant may suggest another mode when your request fits it, but never switches by itself.
• Ask the Assistant about Hive's settings. It explains any setting (what it does, when it helps, where it is in Settings) and suggests one when it spots a problem a setting would fix, such as a very long transcript or agents editing the same files. With Settings → Assistant → Control → Change settings on (off by default), it also makes the change when you ask or agree, and lists it in its panel with the old and new value and a Revert button. Permission modes, the Agent API, what Hive runs and its own Control stay yours.
• The Assistant suggests compacting too. Like an agent's, the Assistant's Compact button and the context in its footer turn orange once its conversation passes a threshold: Settings → Assistant → Highlight Compact over, 500,000 tokens by default (higher than your agents' 200,000, since the Assistant usually works with a bigger context), or Never.
• Start New (All) no longer blanks the Assistant. Starting fresh sessions for many agents at once could turn the Assistant's panel black for a few seconds, after which its terminal drew a little wider than the panel until you resized it. Hive now gives back each ended terminal's graphics context straight away, so the Assistant keeps its own, and a terminal that loses one anyway is fitted to its pane again.
• The Assistant's panel can go on the left. If Windows notifications or other apps cover the right of the screen, View → Move Assistant Panel to the Left (also in the command palette and the panel's ⋯ menu) or Settings → Assistant → Panel side puts it between the project list and your work. It folds and resizes on that side, and the choice is kept for every workspace and after a restart.

Skills

• Agents take turns to merge: the merge slot. When several agents finished close together, one agent's merge could land while another was still running its checks, and that one had to merge main in and run them all again. Now an agent claims the project's merge slot before its final merge (merge main in, check, merge) and releases it after; others wait in order, with "Waiting for the merge slot (B4 is merging #305)" on their status, so each agent's checks run once. Who holds it and who waits shows in the project's Overview and the Progress panel, where Release frees a stuck holder. A hold ends with its agent's session and after 60 minutes without progress (Hive tells you), and if Hive closes while an agent holds it, Hive says so when the workspace opens again. Your merges from the Merge dialog wait for it too. The merge-ready skill uses it, and scripts can read it through the Agent API.
• Agents no longer lose each other's changes to a shared note. Two agents that read a note, edited their own part and wrote it back could each drop the other's change. Reading a note now gives its revision, and an agent rewriting it names that revision: if the note changed meanwhile, the write is refused and the agent reads it again and merges. The workspace-note skill also has agents append dated, signed entries to busy log-like notes instead of rewriting them. Scripts using the Agent API can pass expectedRevision too (Agent API reference).
• Agents merge only what they checked. When several agents merge into the same branch, one could merge work checked against an older main, so main held a combination nobody had tested. The merge-ready skill now has agents note main's commit when they start their checks, compare it just before merging, and if another branch was merged meanwhile, merge main in again and rerun the checks first. It also says to wait and retry when another merge holds git's lock, never deleting it. card-loop points to it.
• A project's Skills tab shows one provider at a time. Pick Claude Code or Codex from a dropdown (it starts on the project's default provider); its skills sit folded under the Hive skills, and Hive remembers, for each project, the provider you picked and what you left open. The list is much shorter.

Hive

• A card loop can't stall unnoticed. On 7 Oct a reviewer's wake line was typed into Codex but never sent, so it sat idle for hours while the builder waited on it. Hive now checks that a wake is taken and presses Enter again if it isn't. An agent in a card loop that ends up idle with no wait gets its card marked as stalled ("Codex isn't watching #333"), and you're told once. The card-loop skill also has agents check that their wait started before they end their turn.
• Long text stays on the card. A card's blocked reason with a long hash or path ran off the card's edge on the board and was cut off. It now wraps, shows its first three lines on the card, and all of it when you hover it or open the card, where the Blocked field wraps too. Long labels, a reviewer's or agent's long name, long titles in the Board's Stalled list and long words in tooltips stay inside their boxes as well.
• Long waits through Hive's tools no longer fail after five minutes. An agent waiting on cards (hive_wait_for_tasks without wake, up to 840 s) or on other agents (hive_wait_for_agents, up to 600 s) got an error after about 300 s instead of its reply. Waits now last as long as asked.
• Hive tells you when files in a project's .hive are already committed to git. Excluding .hive only keeps new files out, so files committed before (with git add -A, say) stayed tracked while Hive reported .hive as excluded. The project's Overview and Project Settings now say how many, with the command that fixes it (git rm -r --cached .hive, then commit), and Untrack… runs it for you after listing the files.
• Tidy the worktrees no agent uses. A worktree you keep when you remove its agent now shows up in the project's Overview under Unused worktrees, with its branch, last commit and size, and whether it is merged into main and clean. Remove deletes a merged, clean one with its branch (or Remove all merged); one that still holds work says what ("3 commits not on main · 2 changed files") and offers Give to an agent… or Remove anyway…, which names exactly what goes. Hive mentions kept worktrees after you remove an agent or delete a template whose agents used them, the Changes tab says when one holds unmerged work, Storage counts their size, and the Assistant can see how many there are (only you can remove them).
• Worktrees reached by another name are recognised. When Hive knew a workspace or a worktree by its short (8.3) name, such as C:\Users\RUNNER~1\…, or through a junction or symbolic link, it thought git didn't list the worktree, because git names worktrees by their full, real path: removing merged worktrees (Remove All, a template's load, Unused worktrees) refused every one, and loading a template again made new worktrees instead of working in its agents' old ones. Hive now compares the real folders; one it can't resolve is still never removed.
• Hive tells you when git is missing. Help → Agent Setup… now has a Git row with the version Hive found, or Git not found with how to fix it, and says what doesn't work with a git older than 2.38. When Hive can't run git (not installed, or not on its PATH), Add Agent, the Changes tab, the Merge dialog, Remove All and template loads say so, and the status bar shows Git in red, instead of saying the project isn't a git repository. A git command that fails no longer counts as "nothing to merge" or clean: Hive keeps any worktree or branch it can't check.
• The card chip takes less room. An agent's pane header now shows only the card's icon and number (#4, +1 with more), as its tab in the agent strip does, at every width; hover it for the titles. The Assistant's workspace list shows the same chip beside each agent, with what it's doing: its test or build progress, background tasks or the cards it's watching.
• The Progress panel keeps a longer history. With many agents, a test run or build scrolled off Recent within minutes. Recent still shows the last 10 runs, and Show all now scrolls through the last 200 in the same space, without moving the rest of the panel; Show fewer folds it back, and each workspace remembers which. The project filter applies to them all, and Recent is still there after Hive restarts.
• Showing an agent no longer takes the keyboard back. After Show the agent in the Progress panel, a key press or click you make before its terminal appears now keeps the keyboard where you put it. Before, the terminal could take it a moment later, so Escape went to the agent instead of closing the run's details.
• Tables work the same everywhere. The keyboard shortcuts, Performance's and Compare's tables, Storage (with Clean Up…'s preview and every project in Settings → Workspace) and the Overviews' tables by project and by agent now sort by any column, filter and page like the compaction history, and remember each table's sort and rows per page.
• Move any dialog out of the way. Like an open card, every dialog (Add Agent, Agent Settings, Merge, questions…) can now be dragged by its header to see what's behind it; it stays inside the window, Escape while dragging puts it back, and it opens in the middle again next time. A question over a dialog moves on its own, and Escape now closes only the question. The image viewers stay put.

Workspaces

• See whether antivirus scanning slows your workspace. On Windows, Settings → Workspace → Antivirus scanning shows whether Microsoft Defender scans the workspace, its worktrees and (for people developing Hive) its test area, and whether each is on a Dev Drive. With your consent, and one administrator prompt, Hive adds exclusions for exactly those folders, or removes the ones it added. Its confirmation lists the folders one per row, long paths wrapping, with a button to copy them all. It suggests this when it matters (several agents at once, or a long build or test run), and if you close the suggestion it reminds you at most once a day, until the folders are excluded or you say Don't Ask Again.
• A deleted worktree comes back. Starting an agent whose worktree folder is gone (deleted after a merge, or not copied to a new computer) offers to recreate the worktree on its branch, with its commits, and runs the project's setup command as for a new one. If the branch is gone too, it offers a new worktree, or to unlink it.
• Move a workspace and keep its agents working. When you open a workspace from a new folder (moved in Explorer, or to another drive), Hive says so in a banner and Repair… fixes what still points to the old folder: agents' worktrees (relinked with git, found whether you moved them along or left them), the folders sessions ran in, and Claude Code's conversations and memory, copied to the folder for the new path so agents keep what they remembered. It shows what it will do first, waits until agents are stopped, never deletes or overwrites anything, and running it again repeats nothing. A worktree it can't find can be recreated on its branch, located or unlinked; a file it couldn't copy is retried until it is. Open Recent lists the new folder, and the projects you were working on stay marked.
• Tidy the recent workspaces. In File → Open Recent and on the welcome page, remove a workspace from the list with its ✕ or right-click → Remove from Recent, or forget them all with Clear Recently Opened… (it asks first, and keeps the workspaces you have open). A workspace whose folder was moved or deleted is greyed with not found, and opening it offers to remove it instead of failing; one open in another window says so.

Coding agents

• See which Claude Code and Codex versions Hive was tested with. Each release now records the CLI versions its tests ran against. Help → Agent Setup shows them next to the versions you have (Tested with 2.1.287 · installed 2.1.290), with a calm note when yours is newer or older, and Copy Diagnostics includes them for bug reports.
• Hive's hook keys stay out of your projects. The key Claude Code's and Codex's hooks use to report an agent's status to Hive was written into each project's .hive launch settings (and into Codex's command line), where a commit, a sync service or a zip could carry it. Hooks now read it from a file in Hive's own data folder, each session has its own key, and a key stops working when its session ends. Each Claude Code session's settings and MCP server configuration, which can carry your MCP servers' keys or a --settings file's env, moved there too, out of the project.
• Hive tells you when a project's .hive isn't kept out of version control. Hive adds .hive to git's exclude file by itself, now also for a repository made later (git init) or one in a folder above the project, and asks git whether it really ignores it. Where it can't (no git, another version control system, a OneDrive or Dropbox folder), the project's Overview and Project Settings say so, with Exclude where Hive can fix it.
• Hive tells you when a sign-in expires. When Claude Code's or Codex's sign-in expired, its agents' turns stopped but Hive kept showing them as working, so card loops sat stuck until someone noticed. Those agents now show Needs sign-in (a violet dot), you're told once for the CLI however many agents stopped, and the tooltip says how to sign in again. Once you have, Resume All Agents in the project header carries on the agents that didn't carry on by themselves. A Claude Code turn that ends on another API error (a rate limit, say) now ends as finished too, instead of staying working.
• A --settings in Claude Code's extra arguments no longer switches off Hive's hooks. Claude Code reads only one --settings, so adding your own replaced the one Hive starts it with: the agent's status, file locks and status line (context, cost, plan limits) stopped working. Hive now merges your settings into its own, so both apply. A --settings with no value, or settings that are missing, aren't a file of at most 2 MiB, aren't JSON, or would turn off Hive's hooks (disableAllHooks), stop the agent from starting, with a message saying which.
• The context tooltip says where Claude Code compacts by itself, and why. "Compacts by itself at about 167,000" is marked as Claude Code's default for the window, and when its settings change it (an auto-compact window in its settings or for one model, --autocompact, CLAUDE_CODE_AUTO_COMPACT_WINDOW, also in a settings file's env block, an earlier percentage, or a --settings file or JSON in an agent's extra arguments) the tooltip gives that figure and names the setting, or says auto-compaction is off. Hive only reads those settings, when the session starts, and only the files Claude Code reads (--setting-sources, --restricted, its 2 MiB limit). A default is called an estimate, since managed policies Hive can't read could change it, and so is any answer when Hive couldn't read a settings file or can't tell which ones Claude Code reads.
• Codex's sandbox setup brings up its prompt on a busy computer too. Agent Setup's Set up for the Codex sandbox could type into Codex while it was still loading, so Codex queued the keys and its setup prompt never appeared. Hive now waits until Codex has loaded.
• Restart session restarts on a busy computer too. Clicking Restart session (after changing an agent's effort, skills or settings) could, on a busy computer, ask whether to stop a session that had already stopped (…already has a running session) instead of restarting it. It now restarts straight away.
• A Codex approval prompt is told once. On a busy computer, Hive could tell you twice that a Codex agent needed your input for the same approval prompt (the first one in a new session). It's told once now.
• Footers show the model that actually runs. An agent or the Assistant set to Opus now reads Opus 5.5, like an agent on the default (Opus 5.5 (default)): the running session's own model, else the model the alias stands for, with chosen as Opus in the tooltip. The model pickers list aliases with their model, Opus (Opus 5.5), as Claude Code reports them.
• New models show up without waiting for a Hive update. Hive now asks Claude Code and Codex which models they have and what each can do, for your installed version and your account (no tokens, no session), so the model pickers list exactly what the CLI reports: Fable, GPT-6.1 Sol, or whatever your plan adds next. Effort pickers offer only the chosen model's levels (Haiku has none, Opus 4.6 stops at Max, GPT-5.5 at Extra high); an effort a model doesn't take stays chosen and is marked, never changed for you. With no effort chosen, the agent's footer shows the model's own default, e.g. Medium (default). The Auto warning for Haiku now comes from what Claude Code says.
• Every fallback is yours to edit. When a CLI can't be asked, Settings → <provider> → Models (fallback) and Effort levels (fallback) are used, and you can add, remove and rename entries, with Reset to defaults. The API prices table now takes new models and removals too, with Reset to defaults, and says when Hive's prices were checked.

Sessions

• One colour per action, everywhere. Stop is red, Remove violet, Start New and New Session green, Archive and Start New teal and Resume amber, in the project header, the agents' panes and menus, the Assistant and the questions that ask before them, so Remove All no longer looks like Stop. Every label is easy to read in both themes (WCAG AA), and labels and icons stay.

• Remove every agent of a project at once. Remove All (n) in the project header (also in the Session menu and the command palette) removes all of the project's agents after one question that lists them, says which are working (they're stopped first) and marks each worktree as merged and clean or kept. Worktrees and branches are kept unless you tick the box to delete the merged, clean ones; unmerged or uncommitted work is never deleted. Open cards are asked about once, and sessions stay in the Sessions tab.

• The Sessions tab is a tree, like the Files tab. Sessions sit under their provider and agent (sessions started outside Hive under Not from an agent), every branch shows its count, and Hive remembers which branches you open for each project. Search filters the tree across names, dates and transcripts and lists each session's matches under it, and the arrow keys move and fold as in Files. Codex's guardian reviews (one per action in Approve for me) now sit under the session that started them instead of flooding the list, and are never offered for resuming. Right-click any branch to Archive All or Delete All at once: Hive asks once, skips sessions that are running or whose files are in use, and says which. Sessions that can't be resumed are marked, with Resume greyed out and the reason on hover.

• Agent templates. Set a project's agents up once and reuse them: Template ▾ → Save Template… on the agent strip keeps each agent's settings and role and the layout, for one project or the whole workspace, and the templates listed under it load into any project, replacing its agents (it lists who goes and who comes first, and waits while an agent runs or has uncommitted work in its worktree). Add Agent ▾ → Add Agent from Template adds just one. Agents also have a Role (Agent Settings), such as builder or reviewer.

• A kept worktree still gets set up. Removing an agent before its first session but keeping its worktree used to skip the project's setup command (e.g. npm install) for the next agent given that worktree. Hive now remembers it, and runs setup for that agent first; a worktree whose setup already ran isn't set up again.

• Edit templates where they are. In the Templates view or a project's Templates tab, Edit… a template to add, copy, reorder or remove its agents, change each in the same Agent Settings dialog as a project's agent (including whether it gets its own worktree), and rename it, describe it or change its layout. Saving checks it first and never overwrites a change made meanwhile.

• Switching templates reuses worktrees instead of piling them up. Loading a template (or adding an agent from one) puts an agent back in the clean worktree of the same name, the replaced agent's or one kept from an earlier load, rather than making hive/claude-2. A worktree with uncommitted work is never reused: a new one is made and the dialog says why. The replaced agents' old worktrees that are merged and clean can be removed, with their branches, in the same step (a tick box, off by default); unmerged or uncommitted ones are always kept, and the dialog says why.

• Loading a template says more before it replaces your agents. For each new agent: its model, effort and mode, and for one with its own worktree, the branch and folder Hive will make for it in this project (never another project's). Agents replaced by a new one with the same name are marked, and the worktree setup command is mentioned when the project has one.

• Worktree agents take less room. An agent's tab and pane header show a worktree icon instead of its branch name; hover it, or Tab to it, for the branch, the folder, where it branched from and what's left to merge. Tabs are narrower and headers keep more of their buttons and card.

• The page buttons sit beside the layouts. On the agent strip, the numbered page buttons now come after Template ▾, right next to the layout buttons they depend on.

• Two buttons for agents and templates instead of four. The agent strip's Add Agent now opens the Add Agent dialog (Configure Agent and Add…), so you choose the new agent's provider and settings first; adding one at once with your defaults is Ctrl+Alt+Shift+N (Add Agent in the command palette). Its ▾ holds Configure Agent and Add… and Add Agent from Template, and Template ▾ holds Save Template… above the templates to load. The menus work from the keyboard (↑ ↓, Enter, → for Add Agent from Template), and Load Template… and Add Agent from Template… are in the command palette.

• Templates view and sharing. A Templates view on the activity bar and a Templates tab in each project list your templates and show what each holds. From there you can load one into any project, rename, duplicate or delete it, and export it as a file to share. Import checks a file before bringing it in, and asks before replacing a template of the same name.

• The compaction history is a table you can work with. In the Overview, a session's compactions come ten to a page (choose how many), can be filtered by trigger or date and sorted by any column, and a click on one opens it in the Sessions tab at its divider, with the summary the agent continued from. Clicking the context in an agent's footer lands on it.

• The context counts the last turn's thinking. An agent's context (footer and Overview) now includes its last turn's output, thinking included, which stays in the conversation and is what Claude Code compacts on. A turn that thinks for 70K tokens no longer leaves the gauge at 116K while the agent compacts at 189K. Hover the context for the two parts and where Claude Code compacts by itself (about 167K of a 200K window); the compaction history says when one turn's output set off a compaction.

• Click an agent's cost to see its session. The cost in an agent's footer now opens the Overview at that agent's current session, as its tooltip says.

• Dragging an agent onto another's pane swaps them. Before, Move here often changed nothing, once agents had been clicked into panes. Now the two agents trade places, running or not, also across pages: hold a dragged agent over a page button for a moment and that page shows, so you can drop it on a pane there. Dragging along the strip still inserts between tabs, and a drop on an empty pane moves the agent there. The panes always show the agents in their order.

• Every agent on a page has a pane. A project now has one layout, and a page holds as many agents as it has panes: four agents in three columns are two pages (three, then the fourth), six in a grid of four are two. Before, a page always held six agents, so a layout with fewer panes hid some of them, reachable only by clicking their tab. One at a time has no page buttons, as the agent tabs already choose who's shown. A project that had a layout per page keeps page 1's.

• Fresh sessions for every agent at once. Start New in the project header (also in the Session menu and the command palette) starts a new session for every agent after one question that lists them and says which are in the middle of something; running agents are stopped first, and their conversations stay resumable. Archive and Start New archives each agent's session first. The header's buttons for several agents say how many they act on, as Stop (3), Start New (6), Archive and Start New (5) and Resume All Agents (2) (the count stays beside the icon when the window is narrow, and follows agents as they start and stop); Stop All Agents is now Stop (n), Explorer and Terminal moved into the header's ⋯ menu, and thin lines now separate the agent tabs. The Project menu now lists every tab, Performance included, in the order of the tab strip.

• Dates as 2026-10-04, or as you like. Sessions named by when they started now read 2026-10-04 14:05 instead of 3 Oct 14:05, in the Sessions tab, the Assistant's conversations, tooltips and exports. Settings → General → Date format (yyyy-mm-dd, dd/mm/yyyy, mm/dd/yyyy or System) and Time format (24-hour or 12-hour) change every date Hive shows. Names you gave sessions are kept as they are, and "2 hours ago" stays, with the date on hover.

• An agent's footer keeps its cost in view. The session it's running is now just an icon (hover for its name and when it started; click and right-click as before). In a narrow pane the transcript size, then the context's tokens, then the whole context show just their icons, then the permission mode's name shortens, and only in a very narrow pane is the cost cut off. Before, the mode kept its full name while the cost was pushed off the edge.

• The context in an agent's footer says what a click does: its tooltip ends with Click to view compaction history.

• See how far along long runs are: the Progress panel. When agents run tests or builds that report progress, a panel on the right, beside the Assistant's, shows each run: which agent, how far along it is ("4 of 12: carddialog"), how long it has taken and about how long is left. Folded, it is a strip with a small bar per run, and it never opens by itself; Ctrl+Alt+P shows it. A passed run fades after a few seconds, a failed one stays until you have seen it and then moves to Recent, and one that stops reporting says so. Hive's taskbar button fills as runs go, and turns red when one fails, so you can see it from another app. Turn it off in Settings → General → Progress panel. The Assistant can see an agent's progress too.

• The Progress panel: details for every run, and failures that tidy themselves away. Click a run, or one under Recent, for everything Hive has for it: the command (copyable), who ran it, when it started and ended, how long it took against what was expected, its steps, its exit code, the log or run record it wrote, and a failure's summary (copyable). A failed run now stays in the list until you've seen it and then moves to Recent, marked red, instead of waiting for you to dismiss it; the strip's and the taskbar's red clear when you see it. Long commands no longer run past the run's box.

• Showing a run's agent says so. Clicking the agent's name on a run now flashes its pane and puts the keyboard in its terminal; before, showing an agent that was already on screen seemed to do nothing.

• Show progress for any long command. Agents run their tests, builds and other commands that take more than about 30 seconds as hive-progress -- <command> (for example hive-progress -- npm test) without being asked, and the Progress panel shows it with the agent's name, the time it usually takes and, when the command prints step lines (##hive-progress step=4 total=12 name=carddialog), its steps, even when it only knows how many after a slow start. The command's output and exit code are unchanged, and outside Hive it just runs the command. It's on the path of every session Hive starts, in Claude Code's and Codex's shells. To have agents use it only when you ask, turn off Settings → General → Agents show long commands in the Progress panel.

• For scripts and tools: POST /v1/progress, PATCH /v1/progress/{id} (which can also give a run started without a total its total, once) and POST /v1/progress/{id}/finish report a run, attributed to the calling agent by its token, and agents' status has a progress field.

• The Sessions tab says where each session ran and which agent ran it: Coder · Project folder, Reviewer · Worktree · hive/reviewer, on every session and in a Ran in line over its transcript (hover for the full folder). A session keeps its own agent: one you removed shows as Coder (removed), and an agent added later with the same name or worktree no longer takes over its old sessions' label. Resume in ▾ shows where each agent works, running ones too.

• Long transcripts are flagged sooner: at 20 MB instead of 50 MB. Well before 50 MB a long conversation slows down the coding agent and Hive. If you kept the old 50 MB default, it moves to 20 MB; a size you chose yourself, Never and project settings are kept. To go back, set Settings → Sessions → Warn when a transcript is over to 50.

• A Codex agent's command left running is counted however Codex runs it. Codex often runs commands from a small script and prints only part of the result; Hive missed those commands, so the agent looked idle while one ran. Hive now also counts one from the id the script printed on its own, or, for a script that runs a single command, from what Codex records.

• Codex cost estimates for GPT-6.1 Sol, Codex's default model, at OpenAI's published prices, and GPT-6.1 Sol in the Codex model pickers. A session on a model Hive has no price for now shows its cost as unknown ($? in the agent's footer, Unknown on the Overview, saying where to add a price) instead of nothing, and cost totals no longer count such sessions as $0: they show Unknown, or what is known with + ?.

• Claude Code agents show the right permission mode with NO_COLOR set. With NO_COLOR in your environment (or Hive started from a shell that sets it), Claude Code redraws only the changed letters of its footer, so Hive kept showing the old mode after Shift+Tab, and switching the mode from Hive failed. Hive now reads the footer as it appears on the terminal's screen.

• Allow on a Codex file-lock notice always reaches the agent. With File locks: Ask me, clicking Allow while Codex was still replying to the blocked edit did nothing: Codex finished waiting for a go-ahead that never came. Hive now tells it to go ahead as soon as that reply ends.

• Switching a running Claude Code agent's permission mode is more reliable on a busy machine. Switch Now and the mode menu no longer give up when Claude Code is slow to show the new mode.

• See what Hive keeps for a project, and clean it up. Project Settings → Storage shows how much space the project's transcript backups, archive, images and agents' worktrees take, measured in the background; Settings → Workspace shows the workspace's total and its biggest projects. Clean Up… moves images and backups of old archived sessions, and images of deleted sessions, to the Recycle Bin, after showing you exactly which files go. Nothing is cleaned up automatically, running and non-archived sessions are never touched, and what the sessions used still counts in the totals.

• A worktree agent's Merge button is an icon with its count, like Compact and Stop, so the header keeps its labelled buttons in narrower panes. Hover it for what it merges.

• A project's tabs shrink to icons when their names don't fit, so all twelve stay on screen in a narrow window instead of some scrolling out of sight; the tab you're on keeps its name, and hovering one shows its name and shortcut.

Windows

• Notices show in Hive while you use it. Windows notifications always appear at the bottom right, over the Assistant's panel where you type, so while any Hive window is focused an agent finishing or needing you is a banner in that window instead: at the top centre (or another place you choose), gone after 6 seconds for a finished agent, and staying until you handle it for one waiting for you. Click it to go to the project. Settings → Notifications chooses what shows while Hive is focused (Show in Hive, Show nothing or Windows notification), which notices get banners (all workspaces, this workspace or this project), where they go and how long they stay. With Hive in the background, Windows notifications work as before. Only when Hive is in the background is replaced, and every profile starts on Show in Hive.
• Keep a Hive window on top. The pin in the title bar, View → Always on Top or Ctrl+Alt+O keeps the window above your other apps, so a test run or an emulator can work underneath. Each window has its own pin, remembered with its workspace when you reopen it or start Hive again.
• File → Close Window (Ctrl+Shift+W) closes just the current window, as its X does. With several windows open, File → Exit now reads Exit Hive (all windows), and its dialog says it closes them all and lists the agents by workspace (telling apart two with the same folder name by their parent folder), with Close this window only if that's what you meant. Close Workspace says it keeps the window open.

Task board

• Decisions on cards. What you decide about a card now has its own place: Decisions, pinned above the comments in the card, each saying who wrote it down and marked new since start when it came after work began. You, the Assistant or an agent you tell can record one (it is always your decision); only you change or remove one. Agents read a card's decisions when they start and again before Review, and follow them over the description; reviewers fail work that ignores one. A decision recorded mid-card doesn't interrupt the agent: its next reply from a Hive tool flags it, and it checks at its next step.
• Archive a whole column, or the whole board, and undo it. Each column's ⋯ has Archive All in Done (23)…, and the board's ⋯ Archive All Cards…, counted per column. Cards an agent is working on, reviewing or waiting on are skipped (and listed) unless you tick Also archive cards agents are working on, and while a search hides cards only the ones shown are archived. Undo in the toast puts them back in their columns and order; later, Archived → Batch → Unarchive this batch does too. An agent waiting on a card you archive is told it was archived.
• Two new columns: On Hold and Passed. The board now runs On Hold, Todo, Doing, Review, Passed, Done. Park work you don't want picked up yet On Hold: only you and the Assistant put cards there. A card that passes its review goes to Passed (reviewed, not merged yet), so Review holds only what is still being reviewed, and Done now means merged: the agent that built a card moves it there after merging its work. Blocked cards stay where they are, marked as before, and nothing on your board moves when you update.
• Find archived cards in a table. Archived on the board now shows a table instead of a long list: sort by any column, filter by number, title, agent, project, label or the column a card was archived from, page through them, and open one with a click. Unarchive brings a card back to the column it came from, and Unarchive Selected brings back every card you tick, or all that match a filter.
• Collapse columns and fold cards. The « in a column's header folds it sideways to a narrow strip with its name and card count, and the other columns widen to fill the board; drop a card on a strip to put it at the top of that column. Each card folds to one line with its own arrow, and a column's ⋯ collapses or expands all its cards. Hive remembers what you folded for each workspace.
• Two agents can build and review a list of cards in turn, without you passing each card between them: tell one "work through #111 to #117 as the builder" and another to review them. Each waits for the other without running, and Hive wakes it with a line when the card changes. The waiting agent shows Waiting for #112 → Review in its header, with Cancel. After two hours with no change, or when the fifth round on one card fails its review (a round is one build and its review), it asks you what to do; a finding that comes back is marked in the review and counts as another round, and one the builder disputes stops the loop to ask you. The wake line names the reviewer of a verdict ("#112 is in Review: Codex (hive) failed it"), and on another agent's card, one an agent's own work depends on, whose card it is and that Passed isn't merged yet ("#111 (Claude's card) is in Passed: Codex (hive) passed it (Passed isn't merged)"); Done means merged. New card-loop skill.
• For scripts: POST /v1/tasks/wait waits for cards to change (a move, a comment, a review verdict), and agents can be woken by it instead of waiting. Agents with a card watch have the new status watching.
• Move an open card out of the way. Drag a card's dialog by its header to see what's behind it; it stays within the window, and opens in the usual place next time.
• A project's Overview shows its cards at a glance, as the Workspace Overview does for the whole board: how many of the project's cards are in Todo, Doing, Waiting for review and Done, and how many are stalled or blocked, updated as cards change. Click a number to open the project's Tasks tab.
• An agent can't finish another agent's work in progress. A card in Doing with one agent can't be moved to Review or Done by another agent (for example a reviewer whose review ended because the card went back for more work); Hive refuses it and leaves the card as it was. The card's own agent, you, the Assistant and scripts with the workspace token still can.
• Agents review cards without taking them. Ask an agent to review a card and it stays in Review with the agent that did the work, showing Reviewing: Codex on the board and an eye on the reviewer's tab, until the reviewer gives its verdict as a comment. Before, a reviewing agent moved the card to Doing, so the board showed the agent that did the work as busy on it. One agent reviews a card at a time, the review ends by itself if the reviewer stops or the card moves, and the Sessions tab says which cards a session reviewed.
• Agents keep to their own project's cards. An agent can see and change only its project's cards on the board: another project's cards are for that project's agents, the Assistant and you. Before, an agent in one project could list, read, comment on and move any card in the workspace. An agent also can't read or type into another project's agents' conversations any more, where those cards appear as their tasks. Each agent now calls Hive with a token of its own, so it can't get round this by saying it is someone else. The Assistant and your board still see everything.
• Moving a card to another project takes it from its agent, and its history says whose it was. Give it to one of the new project's agents afterwards (the card dialog does this when you change both).
• "Check the latest comment": agents can read just a card's newest comment instead of the whole card, which keeps long cards from filling their context.
• Agents move a card to Doing when they start on it. Hand an agent a card in its conversation ("task #59 please") and it moves the card to Doing before it starts, then to Review when it's done. This includes a card back from Review with follow-up work, which before stayed in Review while the agent changed it. (Start… on the board already moved cards to Doing.)
• More work on a finished card goes back through Doing to Review. Start… now works on a card in Done too: it moves to Doing, and its agent moves it to Review when the new work is done, for you to check again, never straight back to Done. The agent's prompt says the card is back for more work. An agent picking up a card in its conversation does the same.
• An agent that starts on a card takes it. When an agent moves a card into Doing, the card is given to it, also if another agent had it (in Todo, Review or Done, so not being worked on); the history says so. A card already in Doing with another agent stays theirs.
• Moving a card to Doing asks who works on it. Drag a card into Doing, choose Move to → Doing, or change its column in the card dialog, and Hive asks: Nobody yet, Assign an agent (nothing starts), or Assign and start, which gives it to an idle or stopped agent as Start… does. Cancel leaves the card where it was. Ask the Assistant to move a card to Doing and it asks the same, unless you've already said which.
• The Assistant sends follow-up work through the card. Ask the Assistant to have an agent act on a card's review feedback and it starts the card again with a note saying what to do, so the card is in Doing before the agent starts, instead of typing the request to the agent with the card left in Review.
• Agents can move cards to Done. Ask an agent "looks good, move #65 to done" and it does; the Assistant can too. Agents still move finished work to Review by default, for you to check. Every move is in the card's history with who made it, and a card can be moved back out of Done from the board or by asking. Before, only you could move cards into or out of Done.

Skills and agent guidance

• A warning over Hive's skills. The Skills view, and a Hive skill while you edit it, say that editing these skills may change how agents behave in Hive.
• Hive keeps its skills up to date in your workspaces. When a new version of Hive improves one of its skills, a copy you haven't changed is updated when the workspace opens, and skills new in that version are added. Before, only a new workspace got them, and existing ones had to Revert each by hand. A skill you edited stays as you wrote it: its page says Update available when Hive has a newer version, and Revert to default takes it. A skill you deleted stays deleted. Hive's personas are kept up to date the same way.
• New skills: work-on-card, coordinate-agents and use-hive-api. Agents carry a card from Doing to Review with work-on-card. The Hive Assistant runs agents and cards with coordinate-agents, its first skill (before, the Assistant got none). use-hive-api helps an agent write scripts against Hive's Agent API. review-agent-work, pick-up, handover, merge-ready, split-work and workspace-note are clearer about when they apply and what they check. For example, a review covers what changed since the last one, and a pick-up doesn't stop to ask when you've said to carry on.
• Leaner instructions. What agents and the Assistant are always told about Hive is much shorter: how to work is in the skills, which load only when they're needed. The rules that matter whatever happens (Review when done, Done only when you ask) stay in every session. The Assistant's personas keep their character, and what it may do comes only from Settings → Assistant → Control.
• Write a skill for the Assistant. Add audience: assistant (or all) under metadata: in a skill's header, and the Assistant gets it. Your other skills go to agents only, as before.
• The Skills view says who gets each skill: Project agents, Assistant or Agents + Assistant, on its page and (when it isn't the default) in the list. A project's Skills tab lists only what its agents get, saying how many Assistant skills it leaves out. Deleting a skill, and Copy to workspace, say who it's for, and what happens to running Claude Code and Codex sessions, instead of "every agent".
• A card back for more work brings its feedback. When a card in Review or Done is started again, the agent's prompt includes the card's latest comment, so the review it came back with is right there.
• For scripts and diagnostics:
  • GET /v1/status says who the caller is to Hive, and which version of the API and of Hive's guidance and skills is running. Each running agent's status says what it launched with, so an agent still on older instructions is easy to spot. Help → Copy Diagnostics includes them.
  • GET /v1/skills/{name} reads one Hive skill, and GET /v1/docs/agent-api serves the API reference for the running version.
  • Agents can read just the newest comments on a long card, page through a long board, and get a short summary of another agent's activity (with the details on request).
  • Skill listings no longer show folder paths, and an agent can't list another project's own skills.
  • A project's skill listing (GET /v1/skills?project=, an agent's hive_list_skills) leaves out the Hive skills for the Assistant alone, as its Skills tab does. An agent's activity says when its task or reply was cut short (clipped), and hive_agent_activity no longer claims detail gives everything: it gives the last ten tool calls and the first 2,000 and 3,000 characters.
  • Polling /v1/status no longer re-reads every skill each time: a skill is checked again only when its files change, and calls at the same time share the work. A skill too big to check (over 64 MB, 2000 files or 12 folders deep) says so, and isn't copied for a new session. A skill's file list says when it was cut short (filesTruncated).
  • Changed for scripts:
    • /v1/status no longer has agent (Claude Code's install info from 0.1): use providers.
    • An agent's activity is now a summary unless you add ?detail=true.
    • Starting a card no longer returns note.
    • workspace.json and project.json drop the skill switches from before 0.2 when next saved.

Performance

• Hive now measures what its own parts cost, for each workspace: Agent API requests, its tools' replies as agents get them, what each session is given at launch, and the skill service's work, next to the providers' reported token usage. Only totals, kept 30 days on this machine, never prompts or replies; no guessed token counts. On by default (Settings → Sessions → Record performance metrics, with Reset); scripts read them from GET /v1/metrics.
• Performance, in the activity bar and as a tab in each project, shows those metrics for the last day, week or month: requests, latency, tool reply sizes, guidance per launch and provider usage, with a trend chart. Filter by project; by who did the work (agents, the Assistant, scripts), which frames the cards, chart and tables it changes; and by provider, next to the launches and provider usage, the only parts recorded per provider, offering every provider whatever is picked. It says how long Hive was recording, so hours it wasn't read as no data rather than zero. Export saves what the page shows as JSON (Shift+click to replace project names).
• Performance → Compare puts a baseline next to a run: views you kept before and after a change, Performance exports, or scenario benchmarks from Hive's development scenarios. Correctness comes first: a smaller result whose checks fail is shown as smaller but failing, and runs that measured different things are marked not comparable, with why.
• Provider usage reads cleanly at any width. Its cells no longer wrap wherever the text happens to break ("≈" on one line and "$25.87" on the next, "Claude Code" in two): Sessions shows the count with 3 running under it, Context the average with 909k max of 1.0M under it, and the cost stays whole, with any unknown sessions under it. Numbers in Hive's tables, and the costs in the Overviews, never split; when a table doesn't fit, it scrolls sideways in its section.
• The Overview charts show every day. In the project Overview and the Workspace Overview, each day of the 7- and 30-day charts is now a tile of its own, as in the Performance chart, so a quiet day reads as its own column instead of empty space. The three charts share one look.

Fixes

• Codex 0.161: switching a Codex agent's mode and Agent Setup's sandbox Set up work again. Codex 0.161 reordered its /permissions menu, so Hive picked the wrong preset: a switch to Ask for approval reported it couldn't confirm, and Set up never reached Codex's sandbox setup. Hive now picks each preset by the name Codex shows, with older and newer Codex alike, and says so if it can't find one instead of choosing another.
• Amber buttons are easier to read in the light theme. The white label on the amber primary buttons (Save, Close, a dialog's default choice) and the amber label of an empty project's Add Agent fell short of the WCAG AA contrast for text. Both now meet it, resting and hovered: the primary fill is the same amber a shade deeper, and Add Agent's label matches Resume's.
• Menus are readable where you point. A hovered or keyboard-chosen menu entry (and the command palette's chosen row) put white text on the light amber, and an entry's second line, such as a session's age in the Resume picker, nearly vanished on the highlight in both themes. Every part of the chosen entry now meets AA contrast.
• The status bar and the count badges are easier to read. In the light theme, the white text on the amber status bar and on the small counts (in the activity bar and beside a project) fell short of the WCAG AA contrast for text, and got fainter on hover; the bar is now a shade deeper amber. In both themes, a status item's secondary part (such as a plan limit) is no longer faded, and the highlighted items (agents needing you, an update downloading or ready) meet AA too.
• Enter on Unarchive unarchives. In the archived cards' table, pressing Enter on a row's Unarchive button opened the card instead; it now unarchives it, as a click does. Enter or Space on the row itself still opens it.
• Filter the Progress panel by project. When runs come from several projects, a dropdown at the top of the panel shows one project's runs (or one agent's in it, or the Assistant's and scripts'), running and recent alike. Each workspace remembers it, and the folded strip follows it.
• Show log in a run's details did nothing for a log outside the workspace, such as an e2e run record. The Progress panel's details now have Open log and Show in folder, which open the log that run reported wherever it is, and say why when they can't. Copy summary is now Copy details: the whole run as text, every field shown, ready to paste to the Assistant. Every run and Recent row ends with a chevron to open and close its details. Agents are also asked to give their runs a short title (hive-progress --title "unit tests" -- <command>), so the panel says what each run is for.
• Agents' output no longer turns into a narrow column after visiting Settings. Opening Settings, another tab or another project while an agent ran could tell its CLI the terminal was a few characters wide; Claude Code then redrew everything one letter per line, and the terminal jumped to the top. Hidden terminals now keep their size, and a terminal scrolled to the bottom stays there.
• The Workspace Overview fits a narrow window. With the sidebar open in a small window (or zoomed in), its period buttons stuck out and the whole page scrolled sideways. They now go under the title and wrap when space is short, and the summary cards stack, while the table by project still scrolls on its own. Even with only a sliver of the window left beside the sidebar and the Assistant, both Overviews now fit: words wrap rather than push the page sideways, and the board's counts, the period buttons and the cards take a row each.
• Dragging an agent only highlights places where dropping it does something. The last agent no longer gets Move here on the spare panes, or a highlighted page button for the page it's already at the end of; dropping there changed nothing.
• A date filter in the Overview's compaction table follows a change of date format. After switching Settings → General → Date format, the When filter kept matching the old format until something else changed; it now matches the dates as they're shown.
• Two windows no longer undo each other's Skills tab and Sessions tree choices. With a workspace in each window, the provider and open groups you chose in one project's Skills tab (or the Sessions branches you opened) could be lost when you changed another project's in the other window, and come back reset after a restart. Each project's choice is now saved on its own.
• The welcome page fits a narrow window. With the sidebar open in a small or zoomed window, a recent workspace with a long path made the page wider than the window, and tabbing to an entry's ✕ scrolled it sideways, cutting off the title and the Start and Recent lists. The two columns now stack when there's no room, and long paths (and, if need be, long names) shorten instead.
• Folders whose names end in .asar work like any other. Since Hive stopped opening a project's .asar archives, it also refused images, PDFs and files in an ordinary folder named, say, art.asar: their previews didn't show and they wouldn't open. Only paths inside a real .asar archive are refused now.
• Closing a window stops the storage measuring it started. Measuring a big worktree for Project Settings → Storage or Settings → Workspace went on in the background after you closed that window (or reloaded it) until it had walked every file. It now stops, unless another page is still waiting for the same measurement.
• Hive no longer closes by itself when you stop an agent and then switch or close the workspace (or quit) straight away. On a busy computer, the agent being stopped could be stopped a second time before it had finished, which crashed Hive, with nothing in its log. Each agent's terminal is now stopped once.
• Escape straight after opening a card no longer asks to discard changes. Pressing Escape the moment a card opened (quicker on a busy computer) could ask Discard unsaved changes? though nothing had changed. It now just closes the card, and every dialog's Escape works from the moment it shows.
• Two windows no longer bring back tips you've seen. With two Hive windows open, what one window's tips remembered (the tips shown, today's tip, commands you've run) could be overwritten by the other window's older copy, so seen tips showed again after a restart. Each change is now saved on its own, and every window follows it.
• A project's Overview fits a very narrow window, session details included. With the sidebar and the Assistant open in a small window, the table under Most recent session (model, version, started, last activity) widened the page so it scrolled sideways. It now scrolls in its own box, its labels whole.
• A window that opens while another changes tips keeps the change. A window still loading could undo what another window had just changed (today's tip shown, so it showed it again; a project's Skills tab or Sessions tree choice), because it applied its older starting copy on top. It now applies such changes again once it has loaded.
• Enter on a recent workspace's ✕ on the welcome page only removes it. It used to open the workspace as well, putting it back at the top of the list. Enter on the entry itself still opens it.
• The image viewer no longer offers Delete for a running session's image. In a project's Images tab and the Assistant's, Delete is greyed out (hover it to see why) while the image's session or conversation runs, as the right-click menu already was; it comes back as you move to another session's image or the session stops.
• Tooltips near the right edge of the window now appear next to their button. The Notifications panel's ✕ showed its "Close" far to the left, over the first notification, and other buttons on the right (the Assistant panel's header, the title bar, the status bar) did the same: every tooltip was placed as if it were the widest one. Each is now placed by its own size, beside its button, and kept inside the window.
• Rebuilding an Electron app in a project could fail while Hive was open ("EBUSY" replacing dist/win-unpacked/resources/app.asar, as npm run dist does). Hive's storage sizes, Files tab and workspace watcher looked at .asar files the way Electron opens its own app, as archives, and kept them open until Hive quit. Hive now treats them as the plain files they are: they no longer stay open, the Files tab lists them as files with their size, and Storage counts their size (it counted them as empty). Leaving a Storage page also stops measuring when nothing else waits for it.
• A project's buttons no longer spill over the side panels when they leave it little room. With the sidebar, the Assistant and the Progress panel open in a small or zoomed window, the empty project page's text and buttons (and a narrow pane's New Session and Resume) spread over the panels and the neighbouring pane. Now everything stays in its own area: in very little room the page keeps just its title and buttons, and scrolls there if even they don't fit. A project without agents shows No agents yet again, instead of an empty pane covering it.
• Your own git add or git commit could fail with "index.lock: File exists" while Hive was open. Hive checks a project's changes and branch status in the background, and git updated its index as it did, locking it for a moment. These background checks no longer write the index, so your git commands (and your agents') don't run into them. Hive's own git actions, such as Merge, still update it as usual.
• Ignored folders such as dist sometimes showed undimmed in the Files tab, when the folder was listed just as another git command (yours or an agent's) was updating the repository. Hive now tries again, keeps the folder's last known dimming meanwhile, and lists it again a moment later.
• A card loop could stall when two cards were reviewed together. If one passed and the other failed a second apart, the builder was told only about the first, and its next wait started after the second, so it waited for a verdict that had already come. Now the line names every card that changed ("#217 is in Review: Codex passed it; #119 is in Review: Codex failed it"), and an agent's next wait counts what others changed since its last wake. The card-loop skill tells agents to check every card they wait on when woken.
• A card loop could stall after a failed review. A builder that fixed the findings and moved the card to Review while it was still there changed nothing, so the reviewer, waiting for the card to come back, was never woken, and both agents waited until you noticed. Now the card's agent moving a failed card to Review where it is returns it for review: its history says Returned for review, round 2, and the reviewer is woken ("#159 (Claudette's card) was returned for review (round 2)"). You can do the same from the board (right-click → Move to → Review (next round)). The card-loop skill also tells builders to move a failed card to Doing while they fix it.
• A skill meant only for the Assistant could go to your project agents instead, when its header wrote the audience in a valid but less common YAML form (a comment after it, metadata: { audience: assistant }, a comment line above it). Headers are now read as YAML. A skill whose header can't be read, or whose audience is misspelt, now goes to nobody and is marked Not given with the reason, instead of going to agents.
• A Hive skill growing while Hive copied it for a Codex session could replace the session's good copy with an oversized one, reported as delivered with no revision. Copies now keep the size limits as they are made and are checked before they go in; a copy that fails leaves the previous one (and says why), for Claude Code and Codex alike.
• Editing a Hive skill or persona while Hive updated it could lose your edit. When a workspace opened, Hive decided an untouched older copy could be updated, and a change you made in the moments before it was replaced was overwritten. Hive now checks the copy again before, during and after replacing it, and keeps your version (and waits for the next time) if it changed.
• A Codex launch could report none of its Hive skills as delivered when one skill's folder had been deleted or couldn't be read at that moment: the agent's details (launched in the Agent API) showed every skill as not copied, though the others were copied and the old copy of the failing one was still there. Each skill now succeeds or fails on its own, and a skill that couldn't be refreshed shows the copy Codex still reads, and why.
• The Hive Assistant on Codex couldn't start when Codex was installed with npm (codex.cmd): Windows ran it through cmd.exe, whose command line is too short for the Assistant's launch, and it stopped at once. Hive now starts a Node CLI's .cmd launcher as its node and script directly, which also passes first tasks and names to it as they are. A launcher Hive can't start directly that would be too long says so and what to do, instead of failing silently.
• The tip card and notifications could cover Resume and New in the bottom-right corner: the Assistant's, or the last agent pane's once its session ended. They now sit left of the Assistant's panel while it is open, and move up above such a bar.
• With a dialog open, the window's minimise, maximise and close buttons stayed bright while the rest of the window dimmed. They now dim with it, as much as the window behind (more with a question over a card), and still work.
• With the command palette open over a dialog, Escape closed the dialog underneath and left the palette open. It now closes the palette.
• The Hive Assistant's Resume and New couldn't be clicked after its conversation ended: the terminal's own drawing layer lay over the bar, so the pointer showed a text cursor there and clicks went to the terminal. The bar is now on top. The Assistant's header also has its own Resume button (|▷) once a conversation has stopped, kept in a narrow panel, so it no longer takes the ⋯ menu.
• In Approve for me, while Codex's reviewer checked an action, the agent's header filled with the whole command (Auto-review: Codex asks to run $env:… npm.cmd run …). The agent now stays Working… with a small shield beside it; hover the shield to see what's being checked. Hive still doesn't ask for you, and agents reading Hive's status see it as reviewing.
• Opening a file, a diff or a Markdown preview with code shrank the icons around the window (the activity bar's from 24 to 16 pixels, and those in the status bar, tabs and buttons) until Hive was restarted. They now keep their size.
• The Notifications bell on the activity bar was easy to miss with unread notifications: only a small dot in a grey icon. It now turns orange as well, clearly visible in both themes, and screen readers hear how many are unread.
• Turning off Settings → Notifications → Flash the taskbar button when an agent needs input didn't stop a flash that had already started; it went on until you switched to Hive. It now stops at once.
• When several agents finished together, their notification came a few seconds later even if you had turned notifications off or come back to Hive's window meanwhile. It is now left out then; with agents in two windows, only those whose window you aren't looking at are told.
• After Compact, an agent you had already given a new message could show as Ready while it worked on it, when the coding agent didn't report the compaction's end and Hive noticed it later. A new message after a compaction now counts as the compaction being over, and the agent stays working.
• Task board: a card's number (#79) was hard to read, in faint grey on the card. It now has the same contrast as the rest of the card's text, in both themes and with column colours.
• In a narrow agent pane, Merge…'s count spilled out of its button over its neighbour, and a long agent name, branch or card pushed the header's buttons (even ⋯) off its edge. The buttons now always fit: a count widens its button, and the agent's details shorten instead (hover for the full text).
• Compact's spinner in the agent header stood still while the conversation compacted. It now turns until the compaction has finished, in the Assistant's header too.
• Task board: long columns while dragging a card. In a column taller than the window you could only drop a card among the cards on screen. Hold a dragged card near the top or bottom of a column and it now scrolls, faster nearer the edge, until the very top or bottom, also with the mouse held still; the drop marker follows, and the card lands where it shows. Near the board's sides the board scrolls to other columns.
• A tab could show another project's list. With the Skills tab open in two projects, switching to one whose skills couldn't be read showed the other project's skills, labelled as this one's older list; a slow read showed them until it finished, and one that finished late could replace the right list. The Sessions list, Memory, Images and MCP tabs, the Files filter, transcript search, the workspace's Skills and hidden projects did the same. Each now shows only its own project's (or workspace's) results, Loading… until they arrive, and its own error with Retry if they can't be read.
• The tray's Need you list with several windows open put the first window's agents before the others', whoever had waited longest, and with ten of them listed, an agent waiting longer in another window was left out under and n more. The list is now oldest first across every window, and To review the same.
• In the compact Projects sidebar, a project's count of agents needing you was cut off at the top of the first tile (and at the side at larger zoom), and could cover the status dot of the project above. The badges and dots now always show in full.
• Codex agents said they needed you while they were still working. In Approve for me, every action Codex's own reviewer was checking showed the agent as waiting for your input, with a chime and a notification, though nobody had asked you anything; a rejected action could leave it showing as waiting until its turn ended. Hive now goes by Codex itself: an agent needs you only while Codex actually shows you an approval or a question. A question Codex asks while it carries on working shows as has a question for you, and is in the list of agents that need you until you answer it.
• File links in the terminal: a path printed in different case from the file (SRC/APP.TS) opened the Files tab without selecting the file or showing its line. Paths with spaces (quoted, or like C:/ws/my app/src/app.ts) and names with brackets (app/[slug]/page.tsx) are now links too, and a piece of a longer path no longer links to some other file.
• Changing a number setting and then ticking Never straight away could leave the number saved and Never unticked, mostly in Project Settings. A number box now saves one change at a time, and your last choice wins.
• Help → Copy Diagnostics… left some of your own text in the log lines it copied: card titles, what the Hive Assistant did, session names, and the names and folders of workspaces you had opened before. Hive now marks your text as it writes its log and the report leaves it out, including what an agent was started with; lines an older version wrote show only their time and source, and folders that aren't Hive's or the coding agents' own are hidden. The report also says which Windows sandbox Codex uses.
• The Assistant's Resume button (on its start page and when its conversation has ended) is amber, like every other Resume button, instead of grey.

0.3.1 — 2 October 2026

Fixes

• Remove Project refused a project whose agent worktree was saved under a short Windows folder name (such as C:\Users\JOHNSM~1\…), saying git didn't list it as one of the project's worktrees. Add Agent → existing worktree had the same problem. Paths are now compared by their real, full name.

0.3.0 — 2 October 2026

Agent footer

• Context as a percentage of the window: the footer shows 84k · 42% once the CLI has reported the session's context window, so you can see how close it is to needing Compact whether the model has 200K or 1M. In a narrow pane only the percentage shows; the tooltip has both numbers.

Notifications

• The taskbar button shows when agents need you: a badge with the count over Hive's icon, the count before the window title (Alt+Tab), and a flash when an agent starts waiting for your answer while Hive is in the background. Each window shows its own workspace. Both can be turned off in Settings → Notifications.
• Agents finishing together chime once and get one notification: "3 agents finished in hive" naming them, or "3 agents finished" with how many in each project, instead of a chime and a pop-up each. An agent asking for your input is still told at once, on its own.

Sleep and shutdown

• The PC stays awake while agents work. A PC set to sleep after a while idle no longer sleeps mid-task because nobody touched the keyboard; once no agent is working it may sleep again (the screen can still turn off and lock). The status bar says so while it lasts. Settings → General → Keep the PC awake while agents work: When plugged in (default), Always or Never.
• Windows shutting down, restarting or signing you out no longer loses the end of a transcript: Hive backs them up first, without holding up the shutdown.
• After sleep, agents' states and plan usage are read again straight away.

When Hive's window crashes

• A crashed window reloads by itself, and says your agents kept running (they run outside the window, so they always did). Before, the window just went blank, and quitting Hive to fix it stopped the agents. If it crashes again within a minute, Hive asks: Reload, Open Logs or Quit Hive.
• A window that stops responding gets Wait or Reload after a few seconds; the question goes away if it recovers.

Bug reports

• Help → Copy Diagnostics… gathers what a bug report needs (Hive's and Windows' versions, your coding agents and whether they're signed in, counts, the settings that change how Hive behaves, and the end of Hive's log) and shows it before copying. Folders, names, keys and tokens are taken out.

Terminal

• Ctrl+click a file path to open it. File paths in an agent's or the Assistant's terminal (src/app.ts, tests/x.test.ts:42:7, stack traces) are links: Ctrl+click opens the file in the project's Files tab with the cursor on that line, in the agent's worktree for a worktree agent, or in another project of the workspace when the path leads there. Only files that exist are underlined.

Tips

• A tip when Hive starts: once a day, a small card in the bottom-right corner shows something Hive can do, with Try it, Learn more (the user guide at that section) and Next tip. It never gets in the way, and steps aside while you type in a terminal under it.
• Tips at the right moment, once each: the first time a transcript turns amber, the context passes your Compact threshold, you add a second agent or you paste a screenshot. Tips about things you already do are skipped.
• Help → Tips… lists them all, grouped and searchable. Don't show tips on the card, or Settings → General → Show a tip when Hive starts, turns them off.

Agent header

• Archive & New moved from the agent header into its ⋯ menu, as Archive and Start New…, leaving more room for the agent's name, card and session.
• Compact and Stop are small icon buttons, like the Assistant's (hover for what they do).
• The session's name moved to the agent's footer, on the right before the context, and no longer repeats the project and agent: until the session has a name it shows when it started (14:05). Right-click it to Rename…; in a narrow pane it gives way first.
• The latest rename wins: rename a session in Hive or with Claude Code's /rename, and Hive shows whichever you did last, everywhere it lists the session. Before, a name given in Hive hid any later /rename, and resuming put Hive's name back.
• Merge keeps your agent's commits by default. The Merge dialog's default is now Merge (the agent's commits and their messages, plus a merge commit) instead of Squash, and it no longer removes the worktree unless you tick it, so one agent can merge task after task. If you had Squash saved, it's switched to Merge once; choose Squash again in Settings → Agents & Worktrees → Default merge style if you want it.
• Squash without the old commits coming back: a squash that keeps the worktree now moves the agent's branch onto the squash commit (Move the branch afterwards, ticked), so its next merge doesn't conflict with its own earlier work.
• No merging mid-task: Merge… is greyed out while the agent is working, asking you something or waiting on background tasks, since a merge would commit its half-finished files. Before, the dialog opened and only said so after you clicked Merge, and a merge while the agent was asking you something went ahead.
• See when a worktree agent has work to merge: Merge… turns orange with the number of commits not merged into the project folder's branch, and the agent's tab shows it too (↑2). A • means uncommitted files only; hover for the details. It updates when the agent finishes a turn, when you come back to Hive, and after a merge.

Agents that need you

• One list of the agents waiting on you. The status bar says how many agents need you (2 need you): those asking you something, and those that finished while you weren't looking. Click it for the list, oldest first, with how long each has waited; click one to go to it. The tray menu has the same list, and the Projects icon and each project in the sidebar show the count.
• Finished means you haven't seen it. An agent that finishes in another project, on another page of agents or while Hive is in the background now stays marked until its pane is on screen. Before, a finish only counted as unseen while Hive's window was in the background, and opening a project cleared all of its agents.
• Work to review: worktree agents that have finished with work not merged are listed below, with what changed (3 files, +120 −40 · 2 commits) and buttons for Changes and Merge….

Lean replies from Hive's tools

• Hive's tools reply in a few lines, not whole objects, so agents and the Assistant use far fewer tokens. Listing the board gives one line per card (number, title, project, agent, labels, blocked, comments) instead of every card in full: about 7,500 characters for 76 cards, where it was over 300,000 and too big for Claude Code to take. Changing, commenting on or reordering cards confirms what changed and where the card is now, in about 100 characters instead of the whole card or column. Projects, shared notes, new agents and started agents reply in a line or two too.
• Full detail on request: hive_list_tasks with details: true, hive_read_task with history: true (the description and comments are always there), hive_session_usage with days: true. Notes and handovers read as their text, and other structured replies are compact JSON.
• Agent API: replies are unchanged unless you ask for the short ones: GET /v1/tasks?view=short, GET /v1/tasks/{n}?history=false, "reply": "short" on card changes, reorders and comments, GET /v1/projects?view=short and GET /v1/projects/{name}/usage?days=false (AGENT_API.md).

Task board

• A task board for the workspace: cards in Todo, Doing, Review and Done, each for a project, with a description, labels, a blocked reason, the cards it depends on, comments and a history of who changed what. Open it from the activity bar (Ctrl+Shift+J); each project also has a Tasks tab. Drag cards between columns; a Doing card shows what its agent is doing right now.
• Start gives a card to an agent (one that is stopped or idle, a new one, or a new one in its own worktree) with the card as its prompt.
• The Assistant and agents use it too. Agents read cards, add cards for follow-up work and move their own to Review with a summary; the Assistant plans work as cards and starts them on agents. Only you move cards to Done: the Assistant asks first, and agents can't. Archiving and deleting cards are yours too.
• An agent that moves a card into Doing takes it: a card nobody had is given to the agent that moved it (or created it in Doing), so it no longer shows as stalled with no agent. A card someone already has keeps them.
• Stalled cards stand out: a Doing card whose agent was removed or isn't running (or that has no agent) is marked in amber and listed under Stalled in the board's sidebar, and the Assistant points them out. Removing an agent that has open cards asks whether to move them back to Todo or leave them.
• See what each agent is working on: an agent with a card in Doing shows it as a chip on its tab, its pane header and the Overviews' running lists (click to open the card), and in the project's sidebar tooltip. Each session remembers the cards it worked on, shown in the Sessions tab.
• Agents and the Assistant can put cards in order: a column's order is its priority, and when you ask them to prioritise ("put the bugs at the top of Todo") they reorder the board itself, a card at a time (top, bottom, or before another card) or a whole list in one go. Each card they move says so in its history. The order of Done stays yours.
• Colour-coded columns: each column has its own colour on its heading and as a light tint on its cards (Todo slate blue, Doing blue, Review purple, Done green). Pick your own in Settings → Board → Column colours, or turn them off.
• Done cards are archived after 14 days in Done: Settings → Board → Archive Done cards after (0 never).
• Agent API: /v1/tasks and the tasks-changed event; hive tools hive_list_tasks, hive_read_task, hive_create_task, hive_update_task, and hive_start_task for the Assistant.

Workspace Overview

• A Workspace Overview (activity bar, Ctrl+Shift+O): what the whole workspace used in a period (7 days by default), every project and the Assistant together: tokens, API-equivalent cost, sessions and prompts, a daily chart stacked by project, every agent running now, a sortable table by project, each provider's plan limits, and the task board at a glance. It reads no more than a project's Overview: token counts come from the usage cache.

Projects

• Project → Remove Project…: Hide a project (Hive leaves it out until you restore it), Remove it from Hive (the folder stays, with its handovers and cards packed into it, so it can move to another workspace and bring them along), or Delete it (the folder, its worktrees and its handovers to the Recycle Bin). Settings → Workspace lists hidden and removed projects, with Restore.

Agents

• See why an agent couldn't start: when its CLI quits before it has started (an argument or setting it refuses, a broken install), the bar under its terminal turns red with what the CLI said and, where Hive recognises it, what to do, with Retry and Agent Settings…. Its tab and header turn red ("Failed to start"), the attention inbox lists it, and a notification tells you if its pane isn't on screen. Before, the reason was only in the terminal.
• Arrange a project's agents: drag a tab in the agent strip or a pane by its header to move an agent, drop a tab on a page button to move it to that page, or use Move Left / Move Right (agent menu, Ctrl+Alt+Shift+Left/Right). The order is saved with the project, and running agents carry on.
• Resume All Agents in the project header (and the project's right-click menu): resumes every stopped agent's last session in one go, leaving running agents alone. If one can't resume, a notification names it and says why, and the others still resume.
• Agents waiting on background tasks aren't shown as finished. When a Claude Code agent ends its turn while a task it started is still running (a test run, say), it shows as waiting on background tasks ("Waiting on 1 background task", a slow, faint dot) and carries on by itself when the task ends. The chime and the "finished" notification wait until it really has finished. Codex isn't told when its background terminals end, so Codex agents show as finished with the count next to them.
• Settings → Agents & Worktrees → Count background tasks for up to (60 minutes): Hive can't tell a test run from something that never ends, such as a dev server, so it stops counting a task after this long.
• Quitting asks first while an agent waits on background tasks (they would stop), and Quit when agents finish waits for them.
• Long conversations are flagged. Each agent's footer shows the size of its conversation's transcript, which turns amber past Settings → Sessions → Warn when a transcript is over (50 MB), and Hive tells you once. A transcript keeps everything, compacting doesn't shrink it, and a very long one slows down the CLI and Hive. Click the size for Hand Over to….
• Hand Over to… the agent itself, in a new conversation: it writes a handover, and a new conversation of its own carries on from it. Hand Over to… is now available with only one agent.
• Clicking the context count in an agent's footer opens the Overview scrolled to that agent's session.
• Use 200K context (instead of 1M) for Claude Code: off by default in Settings → Claude Code, with Inherit/On/Off in Project Settings, a choice in Add Agent and Agent Settings, and one for the Assistant. Current Claude models have a 1M-token window; this holds sessions to 200K, so a long conversation compacts sooner. The Agent API and hive tools take it as context200k.
• The 1M context checkbox is gone from the model pickers: Fable, Opus 4.7+ and Sonnet 5+ always have the 1M window, so it did nothing for them. For Opus 4.6 or Sonnet 4.6 with 1M, type a custom model ID such as claude-opus-4-6[1m]; models saved with [1m] keep working.

Usage

• The Overview and session lists open faster after a restart: Hive remembers each transcript's token counts on disk, and reads a transcript again only if it changed while Hive was closed. Settings → Sessions → Usage cache size (5,000 transcripts) and Clear the usage cache.

Hive Assistant

• The Assistant uses your agents' model and effort. It used to start lighter (Sonnet, low effort), which made it careless: it could say it would check on an agent later and never do it. Settings left at the old defaults move to the agents'; ones you chose stay.
• It waits for agents that are waiting on background tasks, and won't give them a new task meanwhile (hive_wait_for_agents waits through them unless ignoreBackground is set; hive_prompt_agent refuses them). Its instructions now tell it never to promise a later check without a wait actually running.

Feedback for slow actions

• Actions that take a moment show it and can't be interrupted by mistake: Start (on a card), Remove Project, Save, Delete, Archive and Comment on a card, Compact, Agent Settings, Stop All Agents, and deleting a note or an MCP server show a spinner and what they're doing (Starting…, Deleting…); while they run the dialog can't be closed (Escape, ×, a click outside) and a second click does nothing. If one fails, the dialog stays open with the error, so you can try again. Resume All shows a spinner on its button, and an agent being removed shows one on its tab.

Fixes

• /compact typed into an idle agent left it shown as working after the compaction ended, until its next turn. It now shows Compacting the conversation… while it compacts and Ready afterwards.
• An agent that asked the same question twice, or reported its turn's end twice, now notifies (and chimes) once.
• A failed load says so, with Retry, instead of looking like there's nothing: a transcript search that fails no longer shows No matches, the Skills tab no longer shows no skills, the Workspace Overview no longer spins forever, and the same goes for the Images, Memory and MCP tabs, the Files tab's filter, the workspace's Skills view and the hidden projects in Settings. When a refresh fails, the last results stay with a note saying when they're from. In Add Agent, a git error no longer reads as Needs a git repository.
• Number settings: clearing a box such as Suggest compacting above and clicking away saved 0, which turned the feature off without a word. A cleared box now keeps the saved value (in Project Settings it inherits, as before). A number out of range, or text that isn't a number, shows the range under the box instead of a passing message. Settings that 0 turns off (Suggest compacting above, Warn when a transcript is over, Archive Done cards after, Pause after you type) have a Never (or No pause) checkbox; untick it to get your last number back.
• Agent Settings and Add Agent: a long model, effort or permission mode name no longer pushes the boxes past the dialog's right edge.
• Start… on a card you've edited used the card as last saved, so the agent got the old description. It now reads Save and Start… and saves your edits first; if the save fails, nothing starts and the dialog says why.
• Closing a card with unsaved changes (Escape, ×, a click outside, Cancel) lost them without a word, including a comment being written. Hive now asks Keep Editing or Discard; a card with nothing changed closes at once. Save also posts a comment you were writing.
• A question asked from inside a dialog (Delete on a card, say) could appear underneath it; questions now always come on top, and Escape closes only the top one.
• The Changes tab could show the previous file's diff after you picked another while it loaded, and a git error left it stuck loading or showing another folder's files. It now always shows the selected file's diff, and a failure shows the error with Retry; when refreshing a diff already shown fails, the last diff stays with a warning above it that it may be out of date.
• After switching an agent to another conversation, its footer could show the previous conversation's context and cost until the new numbers arrived, or for good if reading them failed. It now shows a placeholder until the conversation's own usage is read, and a refresh that fails keeps that conversation's numbers, faded.
• Merge refuses when git can't check a worktree for uncommitted changes (a damaged index, say), instead of merging and, with clean-up, removing the worktree with them.
• The Changes tab no longer shows the contents of a file outside the project reached through a link (a link is shown as the path it points to, as git stores it).
• Agent Setup → Codex → Set up could sometimes not bring up Codex's sandbox prompt: Hive typed /permissions while Codex was still busy starting, and Codex queued it. Hive now waits until Codex is idle.
• Hive noticed what a running Codex agent wrote to its rollout (its usage, cost and settings) only about once a minute and at the end of each turn: Windows reported the file's old modified time while Codex kept appending, so Hive now checks the size too.

0.2.0 — 1 October 2026

Codex, and a choice of coding agents

• Codex (OpenAI) runs in Hive alongside Claude Code. Each agent chooses its provider, so a project can mix them, for example a Claude Code agent writing code while a Codex agent reviews. Hive shows each CLI's own terminal.
• Settings → Providers turns providers on and off and sets the default provider. New installs start with every provider off, with a banner linking there; updating from 0.1 keeps Claude Code on. Turning a provider off while its agents run asks whether to stop them.
• One settings page per provider (path, default model, effort and permission mode, extra arguments, update checks, API prices), with overrides per provider in Project Settings.
• Help → Agent Setup… has a tab per provider for install, sign-in, updates, and Codex's one-time Windows sandbox setup, which Hive starts for you. It explains the choice between Codex's default sandbox and its non-admin one, and offers Upgrade from the non-admin one later.
• Codex's permission presets: Read only, Ask for approval, Approve for me (the default) and Full access (off unless enabled, like Bypass). Presets and Plan mode switch live in a running Codex session.
• File locks work for Codex agents; with Ask me, Hive asks you in a notification (Allow) because Codex can't show its own approval for it.
• Workspace skills, MCP servers and Hive's own tools reach Codex agents too; file locks, status, notifications, compaction, transcripts, backups and the Sessions tab work for both.
• The status bar has a plan-usage item per provider, and provider icons show on agent tabs, panes and sessions.

Agents

• All agents are equal, and a project starts with none. Add Agent adds one in a click (your default provider, its default settings, in the project folder); its ▾ opens Add Agent… to choose the provider, a worktree and settings. New Session and Resume add an agent when a project has none. Any agent can work in a worktree, and any can be removed once stopped.
• Up to twelve agents per project, six to a page: a seventh opens page 2, and the 1 · 2 buttons (or Ctrl+Alt+PageDown / PageUp) switch pages. A new 3×2 grid shows six at once.
• The layout follows as you add agents: two columns for two, three for three, a grid for four, the 3×2 grid for five or six. Each page has its own layout, and one you choose stays as agents are added.
• Updating from 0.1 clears every project's agents (and resets the layout), since there is no longer a built-in Agent 1. Your sessions stay in the Sessions tab and can be resumed by an agent you add again.
• The Overview's periods count only what happened in them: Today is since midnight, and a session running for days adds only today's work to Today (and each day's cost: its share of what Claude Code reported). 7 and 30 days show a chart of tokens per day. The Assistant's Used so far has the same periods, and the Agent API's usage includes each day.
• The Overview's session details have an agent picker (they followed only the focused agent), and a long compaction history scrolls in its own pane, newest first.
• Settings → Assistant → Pause after you type (15 seconds) and Enter ends the pause: how long after you type in an agent's terminal the Assistant leaves it alone.
• Hive reads only what's new in a running session's transcript when it updates usage, instead of the whole file each time.
• When resuming would re-cache a large conversation, the warning offers Archive and Start Fresh next to Resume.
• Sessions can be deleted in the Sessions tab (and the Assistant's conversations), when they aren't running. Hive's copies go to the Recycle Bin; the CLI keeps its own. What a deleted session used still counts in the Overview.
• Lists show their delete button on hover (skills, MCP servers, shared notes, personas, sessions), and an open item has one in its top bar. Rename now uses a tag icon instead of the pencil.
• Starting a new conversation inside the CLI (Claude Code's /clear or /resume, Codex's /new) is now followed: Hive records and backs up the new conversation instead of carrying on with the old one.

Skills

• Simpler skills: no more switches. Every Hive skill in the workspace reaches every agent in every project. Skills that were turned off in the workspace or a project are on again after updating.
• The Skills view shows the workspace's Hive skills, where you add (+, or Add Skill from File with a .md or a .zip), edit and delete them.
• A project's Skills tab lists everything its agents get: the Hive skills (with Edit in workspace), then per provider the project's own Local (User Managed) skills, which you can now add, edit and delete there, and your user and plugin skills (view only). A local skill can be added for Claude Code and Codex at once.
• Six skills come with Hive, for working with several agents and sessions: handover, pick-up, merge-ready, review-agent-work, split-work and workspace-note. New workspaces start with them. In an existing workspace they're listed greyed out: Restore adds one. When a later Hive improves them, Revert to default on a skill's page brings your copy up to date (the old copy goes to the Recycle Bin).

Hive Assistant

• The Assistant can run your agents. Ask it to add agents, start them on tasks, give idle ones more work, hand one agent's work over to another, change their settings, stop them, or create a project. Settings → Assistant → Control decides how far it may go: Look and advise, Control agents, or Control agents and create projects (the default). It never removes agents or deletes anything, never types into an agent that's busy, asking you something or that you just typed in, asks you before stopping a busy agent, makes at most 30 changes per message, and lists what it did in its panel. Your view stays where it is.
• An agent whose CLI asks whether to trust a new folder now shows as waiting for you, instead of starting forever.
• Each workspace has a Hive Assistant, its overseer, in a panel on the right (Ctrl+Alt+I, or click the Hive Assistant strip down the right edge). Ask it what the agents are doing, about any project, or for a plan or a review. It reads every project and uses Hive's own tools. For now it only looks and advises. It doesn't use a project's agent slot, and hiding the panel doesn't stop it.
• The panel has everything about it:
  • its status and persona;
  • Start, Compact, Stop, and its past conversations;
  • the workspace at a glance, with agents waiting for you first (click one to go to it);
  • its terminal;
  • a footer with its model, mode, context and cost.
• The Hive Assistant view (activity bar) shows what the Assistant has used so far, all its conversations (searchable, like a project's Sessions tab) and its personas.
• Personas decide who the Assistant is, and you can edit them or write your own in the Assistant view. Four come with Hive, each with a serious job and an unexpected character:
  • 🗼 Overseer, a lighthouse keeper keeping a watch log;
  • 🎩 Planner, who plans every task like a heist;
  • 🦎 Reviewer, who reviews code like a wildlife documentary narrator;
  • 🛫 Orchestrator, an air traffic controller.
• Settings → Assistant sets its provider, default persona, and each provider's model, effort, mode and arguments. It starts lighter than your agents, on Claude Code's Sonnet at low effort. Assistant Settings in the panel changes them for one workspace.

Several windows

• File → New Window (Ctrl+K Ctrl+N) opens another Hive window, like VS Code, so you can work in several workspaces at once. Each window has its own projects and agents; settings, the tray and updates are shared.
• Opening a workspace that's already open in another window brings that window forward. Closing a window stops its workspace's agents, asking first as quitting does; so do Close Workspace and opening another workspace in the window, which used to refuse while agents were running. Hive reopens the windows that were open when it quit.
• Agent API: X-Hive-Workspace or ?workspace= names the workspace a request is for, GET /v1/workspaces lists them, and a project can be named <workspace>/<project>. The hive tools always use their session's workspace.

A tidier project view

• Every agent has its own header and footer, with one agent or several. The header has its status, session and buttons (Compact, a red Stop, Archive & New, or Resume and New Session), which turn into icons and then fold into ⋯ as the pane narrows. The footer has its model and effort, permission mode, context and cost.
• The project header is about the project: Active, Explorer, Terminal and a new Stop All Agents, which lists the agents it will stop and asks first. The status bar keeps app-wide items only.
• A narrow window no longer pushes the right side of the project view off screen.
• Running Claude Code agents show their context against the model's window ("39,353 tokens of 1,000,000") in the footer tooltip and the Overview tab, as Codex agents already did.
• Right-click in a Codex terminal pastes once. Codex also pastes on right-click, and it was being sent the click as well as Hive's paste; right-click in Hive's terminals is now always Hive's copy or paste.
• Pressing ← on an empty Claude Code prompt no longer moves the session out of Hive: Hive turns off Claude Code's agent view in its sessions, which put the session into Claude Code's background service, where Stop couldn't end it and resuming failed. Settings → Claude Code → Allow background sessions turns it back on. If a conversation is in the background anyway, resuming it offers Stop It and Resume.
• The status bar's icons are all drawn in its text colour, Hive's own mark and the providers' included, so the Claude mark no longer disappears on the amber bar.
• Switching a session you haven't typed in yet to Don't ask or Bypass (which restarts it) now starts a new session in that mode, instead of failing to resume a conversation that doesn't exist yet.

Other new features

• Hand Over to… in the agent menu hands an agent's work to another agent, of either provider, through a handover; the Sessions tab links the two sessions. Also POST /v1/projects/{name}/handover in the Agent API.
• A new Overview: a project summary for a chosen period (tokens, API-equivalent cost, sessions, prompts) across providers, the agents running now, a section per provider with its plan limits, and a table by agent.
• Estimated costs for providers that don't report one, from a price table you can edit in each provider's settings.
• Share one AGENTS.md in the Memory tab, so Claude Code and Codex agents read the same project instructions.
• Agent API: provider on agents, sessions, usage and status; providers in /v1/status; HIVE_PROVIDER and HIVE_RUN_ID in session environments. Calls that name no agent use the project's only agent; with several they answer 400, with none 409.
• Codex models come from Codex itself, so new ones appear in the model picker without a Hive update.
• The transcript viewer loads the latest messages first and earlier ones as you scroll up, so long sessions open quickly. A running session updates only when you switch on Follow or click Refresh (Settings → Sessions sets the default).
• Overview updates can be live (at most every 15 seconds, the default), every minute, or only on Refresh (Settings → Sessions).
• In the dark theme, the outlines of fields, cards and dialog options are lighter, so they stand out from the background.
• The format help beside an MCP server's definition can be resized by dragging the edge between them.

Reliability and security

• Choosing Haiku with Auto mode (for the Assistant or an agent) warns that Claude Code may run it in Manual instead.
• Switching a Claude Code agent to Manual with Shift+Tab shows in Hive again (Claude Code 2.1.286 shows Manual without its usual hint), and switching to Manual from Hive no longer offers a restart instead.
• A Claude Code session's cost no longer stands still for hours: Claude Code records it only now and then, so Hive adds its estimate for the work since (shown with ≈). The Overview opens on All time.
• The title bar's bottom line runs all the way under the window buttons.
• Hive keeps a last good copy of its settings and records (.bak). A file that can't be read is set aside and the copy restored, with a notification, instead of Hive starting over.
• Hive's copies of workspace skills for Codex are marked, so a hive-… folder of your own in .agents/skills is never removed.
• "Hand Over to…" waits until the handover has actually been written before the other agent starts.
• A Codex permission change is confirmed by Codex before the badge shows it ("Switching to …"); the earlier method could pick the wrong preset.
• Much less file reading while agents work: session lists read each file once, and Codex sessions are found without re-scanning every time.
• Hive tells you once if a CLI version writes transcripts it doesn't understand, instead of showing zeros.
• The hook token is no longer written to the log. The window refuses web permissions it doesn't need, IPC is accepted only from Hive's own page, and links inside the workspace can't be used to open files outside it.
• Links to sections within the user guide now work in Hive's Docs view.
• Unsaved edits in shared notes, skills, instruction and memory files and MCP servers are kept when you change view or file, as the Files tab's are, and quitting, reloading, closing or switching the workspace asks about them. Saving a file that changed on disk since you opened it (an agent edited it) asks before overwriting it.
• Closing or switching a workspace, stopping an agent, quitting or turning a provider off while an agent is still starting now cancels the start, instead of the CLI starting afterwards.
• Turning a provider off lists the agents running it in every window, not only this one's.
• Two agents adding to the same shared note at once both keep their text.
• The Files tab can't read or write through a link (symlink or junction) that leads outside the project; the link itself can still be renamed, moved or deleted.
• A workspace can't be opened inside, or around, one open in another window. A closed workspace's agent worktrees no longer count as part of the next workspace opened in that window.
• Agent API: a workspace name two open workspaces share is refused (409) with both paths, instead of picking one; name it by its path.
• Resizing a pane in one window no longer resets pane sizes changed in another.
• Transcript backups write only what's new: Hive appends what a session's transcript gained since the last backup, instead of copying the whole file each time (a long session's transcript can pass 100 MB). A failed backup is tried again at the next look.
• Reading a long transcript for the first time takes it in pieces, so it no longer needs its whole size in memory at once; Hive keeps the usage of at most 200 transcripts in memory. Two reads of one transcript at once can no longer count its lines twice.
• An agent's hooks are handled in order, so a turn's end can't overwrite the start of the next one (it could show Finished while the agent worked), nor release the file locks, approvals or questions of the next turn. A file reached through a junction or link is locked as the same file.
• Hive types one prompt at a time into an agent, and stops if the agent stops or restarts meanwhile.
• The tray menu is rebuilt at most twice a second, not with every status update of a working agent. Hive notes in its log when its main process was busy for more than a quarter of a second, to help track down pauses while typing.
• When the Assistant asks to stop a busy agent, your answer stops only the run it asked about, and only if Control still allows it. When the Assistant stops, its token stops working and its open questions are withdrawn. The Agent API's event stream is for Agent API callers only.
• Saving settings or workspace settings twice in quick succession can no longer leave the older version on disk. Changing the Agent API settings quickly restarts its server once at a time.
• Typing while a note, skill, MCP server or file is being saved keeps the newer text as unsaved (even when you undo back to what it was), and closing or switching after Save All stays open if something was edited while it saved. A deleted note, skill, persona or MCP server can't come back through Save All, and a note with unsaved edits asks to save or discard them before it is renamed.
• Deleting a shared note or an MCP server sends it to the Recycle Bin, like files, skills, personas and sessions.
• Handovers say who wrote them: Hive starts every handover an agent writes with its project, author (the agent and its CLI), session and date (your time, with UTC). Hand Over to… waits for the handover from that agent's own session, so another agent's meanwhile isn't picked up, and tells the next agent exactly which handover to read.
• Two handovers, notes or MCP servers created with the same name at once no longer overwrite each other. A handover written for <workspace>/<project> is named after the project.
• Saving a file in the Files tab checks and writes it under one lock, so two saves at once can't both pass the check.
• One merge at a time per project, and not while the worktree agent is working. Two agents added at once can't get the same name or worktree, and a new worktree made for an agent that couldn't be added is removed.
• Codex agents in one folder starting together copy the workspace skills one at a time.
• If Hive can't record a session it has just started (a full disk, say), the agent keeps running and is still tracked.
• Two windows opening the same workspace at the same moment can't both have it.
• While the Files tab watches a busy project (a build), it still updates at least once a second, and changes inside node_modules are no longer followed. Moving several files where one name is taken in the destination now moves none of them. A project's Overview doesn't update while another view is shown, and a slow load can't show the previous project's sessions or MCP server.
• Dialogs asked for while one is open wait their turn instead of replacing it. A dialog gives the keyboard back to where it was when it closes; the hidden Assistant strip works from the keyboard.
• Removing the New Session shortcut no longer breaks the empty Session tab.
• The Assistant's start screen and the user guide no longer say it can only look and advise. The Sessions tab's + comes before Refresh, as in the other lists.
• npm run release refuses to upload to a release that is already published.
• Electron 44.5.1.

Notes

• The end-to-end test suites are now in the repository (tests/e2e), with a lint step and a GitHub Actions workflow for pull requests and pushes.
• config.json moves to a new format; Hive copies the old one to config.v1-backup.json first and keeps writing the old fields, so 0.1.x can still read it.

0.1.1 — 29 September 2026

Fixes

• Hive's project and session files could be corrupted, and session records or agents lost, when two agents finished or started at the same moment. Writes are now atomic per write and changes to the same file are made one at a time.
• A Markdown file with a % in an image or link path (such as 50%.png) blanked the whole window. Such paths now work, and a problem in one tab now shows a message in that tab (Try Again, Open Logs) instead of blanking Hive; sessions keep running.
• An agent could be told about another project's handover when one project's name begins another's (hive and hive-website).
• Starting the same agent twice at once (a double click, or the UI and the Agent API together) could leave a Claude Code process Hive no longer tracked. The second start is now refused.
• Archiving a session could keep an older copy of its transcript instead of the latest.
• The Agent API refuses session IDs that aren't IDs, and reports conflicts (agent already running, conversation open elsewhere) as 409 instead of 500.
• Hive only opens the Claude Code files it shows (your CLAUDE.md, auto memory and skills), never credentials or settings.
• Session names containing ", % or & no longer break launching when Claude Code was installed with npm.
• MCP servers imported from a project no longer record the project's full local path in the committed definition.
• The Sessions tab could show messages twice when a running session's transcript was read by two things at once (the live update and a search, say).
• Archive & New no longer starts a new session when archiving failed.
• Plan usage warnings respect Settings → Notifications → Desktop notifications.
• A new Hive skill whose description contains : or # no longer breaks its front matter, and copying a skill to the workspace checks its folder name.
• Unsaved changes in the Files tab are no longer lost without warning: quitting lists the files and saves or discards them as you choose, reloading or switching workspace offers to save first, deleting warns, and renaming or moving a file keeps its changes.
• Far fewer disk writes while agents work: plan usage is saved only when it changes (at most once a minute), and transcript backups are copied at most once a minute during a turn, plus at the end of each turn.

0.1.0 — 29 September 2026

The first public release of Hive. Everything below is what this version does; later releases list what changed.

Workspaces and projects

• Open or create a workspace; every subfolder is a project.
• .hive folders for workspace data (shared notes, skills, MCP servers — meant to be committed) and project data (settings, sessions, backups — excluded from git automatically).
• Mark projects as "working on" to run sessions and get status, without noise from the rest.
• Compact project list: collapse the Projects sidebar to a rail of initials and status dots (Ctrl+Alt+B, the chevron, or drag the edge narrow).

Sessions

• One Claude Code session per project, several running at once, each in its own terminal.
• Live status from Claude Code hooks: ready, working, waiting for input, finished.
• New, resume, stop, archive, rename and adopt sessions started outside Hive. Stop is tinted red and Resume amber in the project header.
• Resume per agent: Resume opens the agent's own last session, and ▾ picks any recent session from its folder; a conversation can only be open in one agent at a time. Each pane shows the name of the session its agent is running.
• Up to four agents per project, each in the project folder or in its own git worktree on its own branch, with their own name and optionally their own model, effort and permission mode. Show them one at a time, in two or three columns, or in a grid; one status dot per project in the sidebar.
• File locks between agents sharing a folder: an agent that tries to edit a file another agent is editing is told to wait or work on something else (or you're asked, or it's warned — your choice).
• Worktrees: Hive creates them next to the workspace, copies .env files and runs an optional setup command; review a worktree agent's changes in the Changes and Files tabs and Merge them back (squash or merge, conflicts detected before anything changes).
• Switch permission mode without restarting: click the mode badge (or Ctrl+Alt+M) and pick one; Hive switches the running session live, and shows the real mode even when you press Shift+Tab in the terminal. Changing the setting offers to switch running agents.
• Compact button next to Stop: summarises the conversation so later messages are cheaper, with an optional focus for what to keep. It turns orange, as does the context count in the status bar, when the context passes Suggest compacting above (200,000 tokens by default).
• Paste screenshots into a session with Ctrl+V, or drag files from Explorer onto the terminal to paste their paths. Images are kept in the project's .hive/images folder.
• Transcript backups so sessions survive Claude Code's cleanup; archived sessions are never deleted.

Reading and reviewing

• Sessions tab: every session of the project with a transcript viewer — the whole conversation including what came before each compaction, thinking and tool calls folded to one line, image thumbnails, live updates for the running session, search across one or all transcripts, copy, export as Markdown and resume.
• Files tab: a file browser (create, rename, delete to the Recycle Bin, cut/copy/paste, drag to move, find by name, git status colours, live updates) with an editor and previews for Markdown (with a split view), CSV/TSV, HTML, SVG, images and PDFs.
• Images tab: every image sent to the project's sessions, grouped by session.
• Changes tab with side-by-side git diffs; Memory tab for CLAUDE.md and Claude Code's auto memory.
• The lists in the Sessions, Files, Changes and Memory tabs and the Docs view, and the Markdown split view, can be resized by dragging the divider; double-click it to reset.

Models and usage

• Model setting (global and per project) with the latest of each family, pinned versions, older versions, a custom model ID and 1M context where the model has one.
• Effort, permission mode, chime and extra arguments per project, each inheriting global defaults. Sessions start in Auto permission mode by default. Bypass permissions sits behind an explicit opt-in. Project Settings has categories and search, like Settings.
• Token use, cache state, re-cache estimate, compaction history and API-equivalent cost per session.
• Plan usage: the status bar and Overview show how much of your subscription's 5-hour and weekly limits is used, with a notification at 80% and 95% of each. The numbers come from Claude Code; Hive doesn't use your credentials.

Skills and MCP servers

• Hive skills and MCP servers deployed in the workspace, enabled globally and turned off per project.
• Machine, plugin and local Claude skills listed for reference, with "Copy to workspace".
• Project .mcp.json servers detected and kept disabled until copied to the workspace.
• Secret detection for MCP definitions.

Agent API

• Local HTTP API with bearer-token auth and a server-sent event stream.
• Built-in hive MCP server giving agents tools for projects, shared notes, handovers (including the latest handover) and notifications.

App

• VS Code-style interface with honey-orange accents, dark and light themes.
• Command palette and Go to Project (Ctrl+P), notification centre, searchable settings with tooltips.
• Configurable keyboard shortcuts for every command (Settings → Keyboard Shortcuts), with per-project shortcuts for project and session commands, and new defaults for agents, layouts, tabs, compacting and the permission mode.
• Completion chime, Windows notifications and system tray with attention badge.
• A quit dialog that lists running sessions and can wait for agents to finish.
• Claude Code setup: detection, one-click install, update and sign-in. The standalone Claude Code CLI is required; editor-extension copies are not used.
• Automatic updates from GitHub releases: checked and downloaded in the background, installed when Hive quits or with Restart and Update, never mid-session. The status bar shows progress; Help → Check for Updates; Settings → Updates to check, download and install manually instead, or to get pre-releases.
• Any number of sessions across projects; only the terminals on screen use the graphics card.
• Open source under the MIT License; the licence and third-party notices are in the Docs view and linked from About.