Every release, newest first

• Workflows are sets of linked tasks that run through the board on their own. Each task has a route of columns, starts when its blockers reach Done, and moves along its route as each column's work finishes. The Workflows view draws the board's cards and their links on a canvas, and a workflow agent plans with you in a side pane. Pick its agent, model and effort in Settings > Workflows. Claude, Codex, Copilot, Cursor, Grok, Kimi and OpenCode can each be the workflow agent.
• Column automations can run when the agent finishes a turn. The On agent turn end trigger pairs with a new Move to column action, a Condition and a Frequency limit, so a column can move a task on when its work is done.
• A Condition can now be a Text match, a pattern checked against the last line of the agent's final message, with * as a wildcard. It is checked in code, so it makes no Automation agent call and costs nothing.
• A completed task opens to a new view. A header strip shows start and end times, elapsed and agent runtime, changes, the pull request, cost and status, over Summary, Usage and Description tabs. Usage lists each run with its columns, model, effort, cost and runtime, every model's tokens, and each background check with whether its automation ran, was skipped or hit an error. The Completed Tasks Duration column is now Agent runtime.
• Background agent calls are now recorded in each project's usage ledger: condition checks, workflow step checks, Knowledge Graph summaries, task titles and Command Terminal names. The usage dashboard shows them in a Background card and a by-day table. Its Cost and Tokens toggle now switches between two separate views, and the token charts read the same per-turn data as the Tokens tile.
• Task titles and Command Terminal names now come from the Automation agent in Settings > Workflows, on its model and effort, instead of the agent CLI's default model. With no Automation agent set, naming is off until you choose one. The Automation agent and Knowledge Graph model pickers now offer the Latest aliases.
• When Claude Code has deleted an idle task's conversation, a resume starts a fresh one from a recap of the task instead of leaving a bare shell, and a toast says so. Kangentic keeps its own archive of each transcript, so the Conversation tab still shows a conversation after the agent deletes its file.
• Every tooltip in the app is now the same tooltip. It opens after a short hover or at once on Tab focus, shows keyboard shortcuts as key chips, and Escape closes the tip without closing the window under it. The context bar's rate limit tip shows each window's share and reset time.
• A paired phone can start and stop Command Terminals. A Command Terminal now ends when its agent exits, WSL terminals included.
• New downloads for Linux on arm64 and for Intel Macs. The Knowledge Graph is unavailable on Intel Macs, because its local model runtime has no build for them.

• Kangentic Mobile is live on iOS and Android. An announcement in the app links to both stores, with a QR code for each.
• From a paired phone you can now pause a running agent and resume a paused one. A phone Resume works like the desktop's Resume button and does not re-run the column's automations. The phone also shows when an agent is starting or resuming, and its session status updates live.
• The Changes panel shows changed images as pictures instead of a binary placeholder. Compare the two versions side by side, with a slider, as an overlay, or as a pixel diff that reports how much changed. An SVG opens on its text diff, and the eye button previews it as an image.
• The context bar's tool-call count and its per-tool table now carry across app restarts and pause and resume, instead of starting again at 0. The Session Summary's By tool table merges every run the same way.
• Dictation has new presets: Best, Balanced and Light, each picking models for your language. Custom adds Parakeet unified, Parakeet v2 and Cohere Transcribe. The Dictation tab lists the running models with their licenses.
• The Knowledge Graph's local model is now IBM's Granite embedding English R2 (Best) or bge-small (Light). An index built on a model that changed is rebuilt in the background after the update, and search answers by keyword until it finishes.

• The Knowledge Graph is a 3D map of your past task conversations, grouped into named regions. Open it from the brain icon in the title bar or with Mod+Shift+A. Filter it by region, time and outcome, and ask it questions. An answer names the tasks it drew on and narrows the map to them. Its first build shows real progress in one card, and task summaries are now written in every project, not only the open one.
• Model pickers lead with a Latest group (Opus, Fable, Sonnet and Haiku for Claude). Pick one and a column, project default or task follows each new release instead of staying pinned to one version. A column pinned to an older version shows a mark in the Column Manager.
• Moving a task to Done, back to To Do, or deleting it now stops the processes its agent left running in the task's folders, such as a detached dev server. A toast reports what was stopped, and its Review list keeps windows and tmux sessions running and lets you stop the rest one at a time. Turn it off with "Stop leftover processes" in Settings > Behavior.
• The tool-calls popover adds Time and Tokens columns, reports correct durations, and opens where it belongs.
• A newer desktop notification for a task replaces the older one in Windows Action Center or macOS Notification Center, instead of stacking beside it.
• Settings > MCP Server lists its tools in collapsible groups by category. The Diagnostics group starts closed.
• Creating a task worktree is about three times faster on Windows.
• The mobile relay now connects through the system proxy, PAC settings and the OS certificate store, so it works behind a corporate proxy or a TLS-inspecting firewall.
• On Linux desktops such as sway, i3 or WSLg with a keyring running, saved credentials are now encrypted and phone pairing works.

• The web demo on kangentic.com shows more of the app. New scenes cover resuming a paused session and dictating into a text field, the Browser scene's page now looks like a real website, and a paused session resumes on the frame it stopped on. Terminals fill their panes at any size instead of leaving an empty band beside or below the output.
• The release poster set is shot in Roboto. The runner used to fall back to Liberation Sans, which has no medium weight and set some labels 17 percent wider than Segoe UI. The zip's manifest now also records each poster's focus area for the docs site.

• Agents can set the browser viewport. kangentic_browser_set_viewport, _pop_out and _dock let an agent check a responsive layout at a real desktop width instead of guessing from CSS. Three mechanisms sit behind them because the wrong one is silently wrong for a given surface: a real window resize for an offscreen or popped-out surface, device emulation for a docked pane. Every path reports the size it measured, never the size that was asked for.
• Seven more browser tools. hover, scroll, select_option, drop_files, history, network and handle_dialog, chosen from an inventory against chrome-devtools-mcp and playwright-mcp.
• One browser surface per task. The isolated option on kangentic_browser_open_pane is gone, along with the per-task lane limit. A task now has exactly one surface: the visible pane, or an offscreen one only when its project is in the background and no pane can mount. That offscreen surface lights the card globe and the Browser pill like any other, the task menu can close it, and opening the pill reclaims it on the page the agent left it on. Before this, an agent could finish a whole verification run in a lane with no browser on screen and nothing to close.
• The drive cue moved onto the page. While an agent drives the browser the terminal is no longer dimmed, because your keystrokes still reach it. The pane is the half that cannot take input, so it carries the veil, a breathing ring, a label and a pointer block, held for the length of the run rather than flickering per burst.
• A GPU crash no longer leaves you with an app that will not start. When Chromium runs out of GPU modes to try, it kills the browser process outright, with no dialog and no warning, and it can do that on every launch: all seven reported runs died within 12 seconds of starting and the user never got in. Every GPU death is now recorded at the moment it happens, and the next launch, finding a fatal one close to the previous run's last sign of life, starts in software rendering instead. That downgrade is a real setting, Settings > Performance > Graphics acceleration, turned off once and never turned back on for you, with a sticky toast saying what happened. Nothing here watches your driver version: we never established what killed the GPU process, so the app claims no cause and suggests no cure.
• A Performance tab in Settings. Graphics acceleration lives there, and Animations moved there from Board, where it never belonged. Memory's own hardware row stays put and is renamed Model acceleration so it reads distinctly from the graphics one.
• The Agent Idle toast does something now. The setting existed, defaulted on, and nothing in the app read it: picking Toast only produced no notification at all, and Both behaved as Desktop only. It raises when an agent finishes its turn or needs permission, and is suppressed when that terminal is already on screen in a task-detail window, the Agent Monitor, or a phone streaming it. That leaves it covering the case the desktop notification skips, where you are here and on this project but not looking at that agent.
• The local model processes leave when idle instead of holding memory all day. Dictation and embeddings each stayed resident for the life of the app once enabled, used or not. On a 9.6 hour session with both idle, dictation held 3.05 GB of Windows commit and embeddings 1.75 GB, together 75% of the app's commit footprint and the reason the low-memory warning fired on a machine with RAM still half free. Embeddings is no longer spawned until something needs it, and opening Quick Find in Smart mode warms it ahead of your first query. Dictation keeps only its small streaming worker warm, loads the larger accurate model on your first press, and recycles back to that baseline when it goes idle. Idle shutdown goes from 5 minutes to 30, shortening to 2 minutes on Windows once a worker passes 1.5 GB of commit. A press starts recording in 6 ms; the first one takes 894 ms to hand back its text while the accurate model loads under your utterance, and 32 ms after that.
• A poster set ships with each release. demo-posters-<version>.zip is attached to the release for the docs site's figures. It is not part of the installed app.

• Columns have automations. A column owns an ordered list of things it does when a task enters or leaves it: send a message to the agent, run a script, call a webhook, or raise a desktop notification. Each row has its own switch, and its own record of what happened the last time it ran. Click a column header to open the Column Manager and build the list.
• A column's message to its agent is one of those automations now. Your existing messages were moved into one automatically, so nothing stops happening and nothing needs re-typing; the field is simply gone from the column's settings card, and the message appears as a row in the list beside anything else you add.
• Template variables paint as you type. Every automation's text fields recognise {{title}}, {{fromColumn}}, {{toColumn}} and the rest, highlighting the ones that will resolve and flagging the ones that will not. Type {{ to pick from the list. A script also receives every variable as a KANGENTIC_* environment variable, which is the quoting-safe way to read one.
• An automation that fails says so. One toast names the automation and the column, with a Run again action; the run log keeps the detail either way. Nothing is retried on its own, because a fired webhook and a half-run script are not safe to repeat blind.
• Three column actions that never did anything are gone, along with the rows that carried them: Kill Session, Create Worktree and Remove Worktree each duplicated something the move already does. If your board still has a Start Agent action carrying a custom prompt, it survives as a legacy row you can edit or delete.
• The Column Manager is easier to use. Remove column is a labeled button in the dialog footer instead of a small glyph on the rail, the automation hint on the rail is a single lightning glyph shown only when something would actually run, and the default height fits every settings row on a tall display without a scrollbar.
• The agent's latest message on task cards. A card's description slot now prints what the agent last said, pushed live as it changes, in the same well the terminal peek uses. Settings > Task > Card Preview picks Latest agent message (the default), Recent agent messages, or Task description. The Agent Monitor card follows the same setting, and the two cards now share one design.
• Goose CLI. Block's Goose is a selectable agent. Resume maps to Goose's own session names and the permission mode maps to its GOOSE_MODE.
• Pick a theme by eye. The Theme tab is a grid of twelve swatches, each painted from its own palette. Hovering a tile previews it on the whole app without committing, arrow keys re-theme live, and a Follow system appearance switch keeps one theme for a light OS and another for a dark one. Two new themes, Rust and Clay, are built from the Kangentic site's tokens; Dark and Light are now named Graphite and Paper.
• Move a task one column without closing its window. Alt+Shift+Left / Right steps the open task while its terminal stays open. Column automations still fire and the existing move confirmations still gate it.
• Images reach the agent as attachments. A pasted or dropped image arrives as a bracketed paste, so Claude Code attaches it directly instead of spending a tool call to read the path. A format the agent cannot take from a path, such as bmp, is re-encoded as PNG first.
• Start a session from your phone. The mobile app can start a task's session again after the agent exited on its own or you paused it on the desktop. A desktop that predates a newer verb now answers with a refusal instead of leaving the phone to time out.
• The Import dialog opens instantly. It paints a per-project cache of the source's items and syncs only what changed since the last fetch in the background. The Open / Closed / All toggle filters the cache without another request.
• Subagent spend in the usage dashboard. The Subagents tile reports nesting depth, and the data groups fan-outs by the turn that started each one, so a task that spawns the same subagent from several turns reads as several fan-outs.
• MCP column tools carry the session fields. kangentic_create_column and kangentic_update_column accept sessionTarget, sessionSpawnStrategy and autoCommandMode, and validate every enum in the handler, so an agent-built Code Review column runs isolated the way a hand-built one does.
• The web demo. The desktop renderer now builds for a plain browser with real recorded sessions, which is what the site and the docs embed. It has no effect on the installed app.

• Merge readiness on the pull request pill. A linked PR now shows whether it is actually ready to merge, not just whether it has conflicts. GitHub counts a merge bypass you personally hold, so a PR waiting on a review you could bypass reads ready rather than blocked, and stays that way once somebody else's PR lands and leaves yours behind the base. Azure DevOps folds its branch policies into the same verdict. Checks that are still running now read queued or running instead of a flat blocked, so the pill tracks CI while it runs.
• Model lists come from the agent CLI. Kangentic asks each agent CLI which models it supports instead of shipping a curated list that goes stale between releases.
• Remotes fetch in the background, and "behind" is measured against the base branch. The branch pill stays current without you running a fetch yourself, and the ahead/behind counts compare against the branch the work actually branched from. The refresh interval is configurable, or can be turned off.
• Moving a task from your phone answers as soon as the move commits, rather than waiting for the whole board to settle.

• A task now tracks the branch it actually pushed. Kangentic records the pushed branch and the base branch it resolved against, so a task whose branch was renamed locally, or pushed under a different name, no longer loses track of itself.
• Pull request linking follows that pushed branch instead of guessing from the task slug. A task whose branch diverged from its slug now finds its PR.
• The commit anchor in the Changes panel appears only when the connected board provider can actually supply it, rather than offering a control that leads nowhere.

• Semantic memory works in packaged builds. The embed worker was forked without part of its dependency closure, so it exited at startup on every installed copy and semantic search never ran. If you have only used a released build, this is the first version where it works at all.

• Azure DevOps pull requests. A task on an Azure DevOps remote now resolves and shows its PR the way a GitHub task does, including a pasted /pullrequest/<id> URL. PR lookups are dispatched only to the connector that owns the remote, so a remote Kangentic cannot read reports a clean miss instead of clearing the task's existing PR link.
• The Branch row says where the task will run. The worktree on/off toggle is now a Worktree / Project pair, so the project folder is a named choice rather than an unnamed off state. When a worktree is impossible (not a git repository, a repository with no commits, or a project folder that is itself a worktree) that option is disabled with the reason as its tooltip, and the hint below states plainly where the agent will run. The task detail menu now names the folder it opens.
• A {{projectPath}} template variable. Auto-command and spawn-agent prompt templates can now reference the main project checkout, which {{worktreePath}} could not express for a task with no worktree. It lets a column hand an agent something like git -C {{projectPath}} merge {{branchName}}.

• The Changes tab is now a proper review surface. It splits into a changed-file rail and a diff pane with a resizable divider, and the rail carries a collapsible History section: a commit graph you can click to scope the diff to any single commit, plus a pinned "Uncommitted changes" row. Mark files viewed as you go, sort and filter the list, right-click for a file's own history, and double-click any file to pop its diff into its own window.
• View options for diffs. Ignore whitespace, collapse unchanged regions, wrap long lines, and render inline when the pane is narrow, all as named checkable items shared with Settings > Changes. Per-file blame annotation lives in the same menu.
• Close a task's browser to free its memory. Hiding the Browser pane keeps the page loaded on purpose, so an agent driving it is never interrupted when you reclaim the space. When you are genuinely finished with a task's browser, "Close browser" ends the page and gives back the memory (roughly 120 MB per loaded pane). The task's URL is remembered either way.
• More reliable browser panes. Panes now resolve by stable surface handles and by their own task, stay alive while hidden, and park correctly when a window closes instead of being dropped.
• Crash and error reporting. Kangentic now reports crashes through Sentry alongside the existing anonymous usage events. Stack traces have machine-specific paths stripped, and no task content, code, or personal data is attached.
• Faster CI. The pipeline consolidates into a single checks job with 11 UI shards, and the code-review pass gathers its review pack once instead of once per reviewer.

• Each task's agent browser gets its own cookie jar, but you sign in once. A project's identity-provider login (Google, Okta, and the like) is shared across that project's tasks, while each task's localhost session stays isolated, so one task's dev-server session can never be mistaken for another's.
• A task never starts from a stale base branch. Kangentic checks that the base is current before cutting the worktree, and the Changes panel offers "Update from base" when a task's branch has fallen behind.
• Pop out a single file's diff. Double-click a file row in Changes to open that one diff in its own window (up to eight at a time), so you can read one file while the panel stays on another.
• Filter the Agent Monitor by project. The toolbar's scope filter now takes a multi-project selection instead of one project at a time.
• Agent task tools answer more up front. The MCP task tools now return four things an agent used to have to discover by probing, so it spends fewer calls working out what it is looking at.
• Dev builds say so. An npm start window names itself "Kangentic (dev)" in the wordmark and the taskbar, and warns loudly when another instance already holds the single-instance lock instead of silently handing you the wrong build.

• Every concurrent agent gets its own browser. Each agent now drives a private offscreen browser lane instead of contending for one shared guest, CDP drives are serialized per guest so two agents can no longer interleave mid-command, and every drive records its caller so contention is visible when it happens. A lane is torn down with the session that owns it, and an agent's browser now survives the task window being closed.
• Dev-server ports are leased per task. A global registry hands out a dev-server port that nothing else is using, so two agents starting servers at the same moment cannot pick the same number, and a task's Browser pane can default to its own lease. kangentic_check_dev_ports reports what a task holds and what is actually listening on the machine, and the default lease range has been moved clear of every common framework default. In prompt templates, {{port}} resolves to the task's own reserved port.
• A browser pane that hits a build error says so. Instead of returning a screenshot of a dev server's error overlay, the agent gets the build error itself, named file and all, with a note that the break may belong to another agent sharing the worktree.
• An agent's focus move is shown rather than hidden. The terminal dims and the pane is marked while an agent drives the browser, so keystrokes never land somewhere surprising. OAuth popups are allowed through.
• Find tasks by ticket number. Search # followed by a task's ticket number to jump straight to it.
• Dictation targets whatever text field has focus, not just the agent terminal.
• Auto-update now works on Linux.
• Clearer board vocabulary. The Tests and Ship It columns are now Testing and Merge, and auto-command is labelled as a message to the agent.
• Spawn failures are no longer silent. Starting a task on a branch that is already checked out is now rejected with a reason instead of failing quietly.

• Two new agent adapters: Grok Build (xAI) and Antigravity. Both ship with the full Claude-parity harness: caller-owned session resume, hook-driven activity detection, Kangentic MCP wiring, folder trust, and usage telemetry. Kangentic now supports 14 agent CLIs.
• Verified auto_command delivery for six more agents. A task's opening prompt is now confirmed delivered rather than fired and forgotten, with a warning toast naming the command and reason when delivery fails.
• In-app announcements got a home. A megaphone in the title bar carries an unread badge and opens a browsable history, so an announcement you dismissed is still readable later.
• Mobile companion pairing is honest in both directions. The desktop announces a revoked pairing, and acts immediately when the phone unpairs from its side.
• The Kangentic MCP server is now wired into Codex, Gemini, and Droid, so agents on those CLIs can read and drive their own board.

• Docking a window follows your cursor - drag a task window by its header and wherever you point decides what happens. Run into the left, right, or bottom edge of the board to snap that half, into the top edge to maximize, or over another window to tile beside it: its left and right thirds dock to that side, the middle third docks above or below. A nudge just repositions, and Escape mid-drag puts the window back where it started. The separate tile-layout menu is gone, since the drag now expresses every layout it offered.
• Click outside a task window to close it - the old behavior only worked when a single window was open. Now the focused window closes on an outside click however many are open, and the first click still lands normally on a control, a task card, or a running terminal, so dismissing never costs you a click you meant for something else. Set it at Settings > Behavior > Close on Outside Click.
• Pausing a session puts its window away - pausing from the task detail header or kebab now closes the window too, instead of leaving you to dismiss it separately. The session stays paused and resumable from the board.
• Agents can open their own Browser pane - an agent that needs to look at your dev server can now open and close the embedded Browser pane for its own task over MCP, rather than stopping to ask you to click the Browser pill. Pane control is scoped to the agent's own project, panes survive a project switch, and the pane list reports each pane's live URL.
• Agents can place and reorder tasks within a column - task position is now something an agent can set, not just which column a task lands in.
• Auto commands are verified, not fire-and-forget - the delivery path for commands attached to a column transition was rebuilt so Kangentic can confirm the agent actually received one, and report it when that fails. Injection now uses Claude Code's documented keys and no longer interrupts an agent mid-turn.
• A more consistent settings and dialog surface - setting labels, descriptions, and input fills now come from shared components, so a dropdown, a text field, and a toggle card read as one family wherever they appear.
• Snappier boards with many terminals - terminal construction is serialized rather than done all at once, and the context bar's fill is composited.

• Mobile companion app, out of dev builds - the mobile bridge now ships in production, so you can pair a phone from Settings > Mobile Devices on a normal install. The relay picker names the hosted relay and backs its "Official" badge with a published page you can check, the pairing steps link to the full docs, and a fresh install gets a "Pair a phone" link on the welcome screen.
• In-app announcements - product announcements now arrive inside Kangentic instead of only on the website, with a "Get the App" section in Mobile Devices.
• Agent Crash notifications in Settings - the Agent Crash notification finally has a control in the Notifications tab. It was previously reachable only by hand-editing config.
• Self-repairing terminal frames - a terminal session that reopens with missing rows now detects it, repairs the replay, and captures what went wrong.

• Release notes on first launch - after an update installs, Kangentic shows what changed the first time you launch the new version, in a What's New dialog with its own icon and title.

• Cross-project agent monitor - one view of every running agent across all your projects, with a live output peek on each card and a shared task detail you can open without leaving it.
• Board profiles - per-column agent, model, and effort presets that a task picks up as it moves across the board.
• First-launch walkthrough - a five-step guided tour that introduces the board, columns, and agents on a fresh install.
• Rewritten description editor - a source-first markdown editor with compact attachment chips, plus a paste-as-plain-text path.
• Command Terminals in the sidebar - live Command Terminals are surfaced per project and can be named, and the sidebar footer gains a split button and a background menu on the project list.
• Release notes before you update - the updater now shows what changed in a modal before it restarts to apply an update.
• Column tools over MCP - agents can create and delete board columns, and column removal can be staged from board settings.
• Send-to-session steering - push a message into a running agent session over MCP, with a durable log of everything that was sent.
• Task-numbered worktrees - worktree directories are named after the task number, so they are recognizable on disk and stay well under Windows path limits.
• Preview exit notifications - the session that launched a preview is told when that preview exits.
• Agent Override editing - an edit pencil on the Agent Override branch of the task dialog, and the task run mode now survives a save.
• Terminal OSC 8 links open in your default browser with no confirmation dialog, and the activity glyphs across the board, sidebar, and cards now come from the shared branding set.
• Faster board and terminal - board hot paths are cut and cards stay live during a drag, terminal initialization is deferred a frame, monitor snapshot pushes are gated, and task transcripts are revalidated by file stat before being re-parsed.

• Terminal color overhaul - a near-black background, a real 16-color ANSI palette, and fully custom background/foreground/cursor colors, including a one-click preset that matches your app theme.
• Task settings tab - task-presentation settings now live in their own tab, and the terminal font is picked from your detected system fonts via an autocomplete.
• Word-delete on Backspace - an opt-in terminal setting so Backspace deletes the previous word instead of a single character.
• Disable Codex's ChatGPT Apps connector - a per-agent launch toggle that skips the optional cloud connector that can hang Codex startup.
• Remote OpenCode servers - point a task's worker at an OpenCode server you run instead of spawning a local process.
• Configurable MCP server bind address - choose the address the built-in MCP HTTP server listens on.
• Better activity tracking - a live idle-wait countdown, a durable history of active/idle intervals, and faster recovery from resume-picker turns.

• Usage statistics dashboard - a new project-level and app-wide view of token usage, cost, and session activity, with live and cumulative breakdowns.
• Mobile companion bridge - pair a mobile device over a secure relay transport to watch and drive your boards, backed by a typed protocol package.
• Detachable panes - pop the Stats, Changes, and Browser surfaces out into their own movable, resizable windows.
• Refreshed branding - new Warm Craft desktop app icons and a theme-adaptive title-bar mark.

• Conversation viewer upgrades - long transcripts now scroll smoothly, with in-viewer search and open-at-position so you can jump straight to the turn you care about.
• Per-project agent defaults - set a default model and reasoning effort per project, override the model, effort, and permission mode per task, and configure MCP servers from agent settings.
• Repo history in the Changes panel - browse the commit history, view any file's history, and run blame without leaving the task.
• Move a task to another project - relocate a task across projects, also available to agents via the kangentic_move_task_to_project MCP tool.

• In-place task description edits via MCP: agents can now update a task's description with targeted find-and-replace edits through kangentic_update_task, instead of resending the whole description.
• Background conversation indexing: a central embedding engine now indexes conversation memory in the background, duty-cycle throttled so it stays out of the way of active work.

• Searchable conversation memory. Semantic + keyword search (hybrid RAG) across your past conversations.
• Description peek. Open a task's full description as a resizable side panel and toggle it with Mod+Shift+K.
• Per-task commit graph. A new commit graph pane in the task detail view.
• Copyable terminal output blocks. Select and copy whole output blocks straight from the terminal.
• Markdown preview in the diff viewer. Toggle a rendered markdown view of changed files.
• Smarter model picker. Rescans installed models when you open the dropdown, hides superseded model generations, and shows a self-discovered context-window badge.
• Redesigned Edit Columns. A master-detail layout with maximize support.
• MCP improvements. Attach and remove files on existing tasks, no more prompts for Kangentic's own MCP tools, and each MCP tool pill now links to its docs.
• Snappier UI. Eliminated a class of recurring UI freezes.

• Push-to-talk dictation. Hold a hotkey to speak and have your words typed straight into the terminal, with local on-device voice-to-text.
• Agent-driven browser pane. Agents can now drive an embedded Browser pane through the new kangentic_browser_* MCP tools, with a streamlined Agent Browser settings tab.
• Ollama adapter. Run local LLMs through Kangentic with the new Ollama agent adapter.
• Team-shared column descriptions. Add a shared description to any board column; it round-trips through kangentic.json for the whole team.
• Ticket numbers on cards. Opt in to show ticket numbers on task cards.
• MCP task-creation cap. Set a limit on how many tasks the MCP server can create, from settings.
• Persistent task layout and stats. Each task remembers its detail-view layout and accumulates lifetime session stats.
• Task title in preview. The preview header now shows the task title.

• Changes panel for branch review. Review your branch diff in-app with a branch header, a segmented diff-scope control, and a resizable auto-fit file tree. Toggle between tree and flat views, sort files, collapse all, mark individual files as "viewed", and navigate across files from the keyboard. Git metadata refreshes automatically as you work.
• Command terminals are now full windows. Drag, resize, snap, and tile command terminals on a dedicated, globally-persisted layer, and run several tiled side by side.

• Modeless task-detail windows. Drag, resize, snap, and tile task windows across the app shell, and click outside or middle-click to close them.
• Universal keyboard and mouse bindings. A unified binding set across the app, including middle-click to close task windows.
• Saved diff view preference. The split/inline git diff view is now a global preference, labeled "Git Diff View".
• Rate-limit reset marker. Rate-limit bars now show a session reset time indicator.
• Activity-aware pause/resume. The pause/resume button reflects the session's live activity state.
• Configurable worktree setup. Worktree creation runs your init script, with configurable node_modules linking plus progress and cancel support.
• Background PR refresh. Pull request state refreshes in the background on project open and on a timer, with a clearer clickable state badge.
• CI-driven test gate. The Tests column opens a PR and the Ship It column merges it once CI is green.

• Cross-agent transcript views. get_transcript now works across every agent with structured output -- responses/result views, tailing, search, and size caps -- via a shared transcript-parsing capability.
• One-step project relocation. Re-point a project at a new folder in a single action: Kangentic moves the folder, quiesces its own running sessions, and migrates per-agent project data for all agents.
• Roomier task dialogs. The New Task and Task Detail dialogs are larger, support maximize/restore, and have cleaner close controls.
• Unified overlay motion. In-app overlays share one consistent animation, with double-click-to-maximize on dialog headers.
• Smarter Changes panel. The diff view centers the first change on open and restores your scroll position when you revisit it.
• Quieter first move out of Done. Moving a task out of Done for the first time no longer re-fires the column's auto-command.

• Relocate a project. Move a project to a new directory without losing its board, tasks, or history.
• Smarter model picker. The model picker auto-discovers newly released models and groups [1m] context variants and dated snapshots in the dropdown.
• Customizable hotkeys. A new Hotkeys settings panel, backed by a central keybinding registry, lets you see and customize keyboard shortcuts.
• Maximize task detail and terminal. Maximize/restore controls and keyboard shortcuts for the task detail view and command terminal.
• Resizable task detail split. Drag the divider between the terminal and the right-hand panel in task detail to resize the split.
• Unified toolbar. The board and backlog now share a single toolbar with shared search and filter.

• Per-column session control. Columns can now run their own isolated agent session, with separate settings for the session target and spawn strategy. These per-column settings round-trip to kangentic.json so they can be shared with your team.
• Live session stats in the context bar. The context bar now surfaces live session stats with an expandable tool-call breakdown.
• Smarter code review. Code review scopes against the base branch through your working tree, and larger reviews automatically fan out across multiple agents.
• More reliable PR linking. Branches now resolve to their pull request through a confidence-ranked resolver.
• Clearer diagnostics. Main-process logs are tagged with the project name and prefixed with local-time timestamps.
• Readable Isolated badge. A theme-adaptive Isolated badge now appears on the session tab and task detail.

• Right-click context menu for archived tasks on the board lets you act on archived items without dragging them out of the archive.

Per-task agent, model, and effort overrides

Pre-spawn picker lets you choose a different agent, model, or effort level for a single task without changing the column or project default. Overrides are remembered per task and persist across resumes via a model cache. Falls through to the column override, then the project default, then the agent's built-in default.

Chrome-style zoom in the embedded browser

Ctrl+Wheel, Ctrl +/-, and a toolbar pill all control browser pane zoom. Zoom level is preserved per task. Ctrl+Enter in the note input is now scoped, so it sends to the agent without stealing focus.

Pasted screenshots auto-compressed

Screenshots pasted from the clipboard are downscaled and re-encoded to fit within Claude's API image budget, preventing oversize-image rejections.

Unified MCP task search

kangentic_search_tasks now searches across the active board and the backlog in a single call, with the source surface returned for each match.

Persistent lifetime usage stats

Lifetime usage stats now survive task and session deletion. The StatusBar period selector (Live/Today/Week/Month/All Time) keeps a complete history regardless of which tasks have been archived or removed.

Activity engine: trace capture, replay, and overlay

The activity debugger gained a trace capture/replay pipeline, a timeline overlay (Ctrl+Shift+D), invariant fuzzing in the test suite, and a memoized grid layout. Background-shell detection now uses snapshot health for probe-failure signaling and preserves named bg shells across HMR.

Devtools surface for agents

MCP devtools surface expanded with screenshots, console, mutations, React tree/recent renders, IPC log, latency telemetry, and a preview-inspection bridge. All dev-only, excluded from production builds.

/test skill maintenance

/test now self-maintains its affected-test selection rather than relying on a manually curated list. /code-review auto-fixes findings by default.

Four new agent adapters

• OpenCode (sst) - hook-based activity stream via plugin, MCP server wiring via OPENCODE_CONFIG_CONTENT, Plan/Build permission modes mapped to OpenCode's --agent flag, auth detection via ~/.local/share/opencode/auth.json, and TUI transcript cleanup for cross-agent handoff.
• Qwen Code (Alibaba) - caller-owned session IDs via --session-id, auto-pre-populated folder trust for spawned worktrees, MCP server wiring, transcript cleanup, and full E2E session-resume coverage.
• Kimi Code (Moonshot AI) - full wire-protocol v1.9 telemetry (context %, token counts, tool events) via ~/.kimi/sessions/<hash>/<id>/wire.jsonl, caller-owned session IDs via --session <uuid>, --continue for resume-latest, subagent lifecycle tracking, PlanDisplay markdown rendering in the Activity log, and CLI auth detection.
• Factory Droid - native session JSONL parsing for the Transcript tab, Ink TUI scrollback cleanup for handoff, and a "no live telemetry" capability flag so the UI sets accurate expectations for Droid's PTY-only model.

Embedded browser pane

A sandboxed Chromium pane lives inside the task dialog. Point it at any URL, draw annotations, pick DOM elements, and submit the rendered frame plus context to the active agent as a multi-modal prompt, all without leaving the task. Includes a clear-data action, a settings tab, and a shared partition constant.

Global search palette (Ctrl+Shift+F)

One overlay searches across active tasks, archived tasks, backlog items, session events from events.jsonl, and registered projects, with per-kind grouped results and keyboard navigation. Selecting a session-event hit scrolls the Activity Log to the matched event with a brief highlight; backlog hits switch to the backlog view and open the item. Also exposed to agents as the kangentic_search_everything MCP tool.

Per-column and per-task model & effort overrides

Pin Plan to opus, Execute to gpt-5-codex, Review to a cheaper model. Kangentic live-applies model and effort changes via the agent's /model and /effort slashes when sessions cross a column boundary, with no respawn needed for adapters that support live swap. Per-task overrides via the new context-bar popover take precedence over the column override; both fall through to the agent default when unset.

Auto-name tasks from the prompt

Task titles can be generated from the description via a new Sparkles button next to the title field, plus a 30-second toast that suggests a rename for placeholder-titled tasks. Implemented per-adapter (Claude, Codex, Gemini, Qwen, Kimi, OpenCode, Cursor, Droid, Copilot) using each agent's plain-text headless mode. Configurable rate limit and a "don't ask again" persistence key.

Activity engine rewrite

The 3-guard state machine was replaced with a single-predicate ActivityEngine with declarative watchdogs and a process-tree-based background-shell watcher. Tier A (PID-aware) plus Tier B (count heuristic) detection of background-shell exits, with cycle-shared OS process queries to fix the 10-session scaling cliff. New IPC paths (session:notifyUserInterrupt, session:getActivityReason, session:getActivityStats) feed a Developer settings tab with a live debug overlay (Ctrl+Shift+D). Fast Ctrl+C recovery via a 3-second user-interrupt coordinator, where the previous floor was 5 minutes through the stuck-pending-tools watchdog.

Other notable additions

• Per-tool stats breakdown captured per session with cost, tokens, durations, and interrupted counts.
• Rate-limit pill in the context bar synced across all active agents, generalized for adapters that report multiple limit windows.
• Claude effort level shown next to model name in the context bar.
• Optimistic delete for tasks with snap-back if the IPC fails.
• Last active task tab remembered per project.
• Analytics events tagged with the agent name and resolved model so launches/exits/completions are attributable.
• Task prompts and handoff context wrapped in a structured XML envelope (<task><title>...</title><description>...</description></task>) with a new {{task_xml}} template variable, replacing ad-hoc string concatenation.
• MCP routing by prompt project cues so cross-project tool calls land on the right project automatically.
• MCP create_task now auto-attaches files referenced by absolute path in the prompt.

• Auto-Resume Agents on Restart toggle in Behavior settings. Decide whether projects should resume previously-running agent sessions on open, or wait for you to click Resume on each task. Useful when resuming many agents at once slows your machine.
• Cross-project MCP tool calls. Pass an optional project parameter to most MCP tools to query or mutate other projects without switching context. A new kangentic_list_projects tool helps agents discover available projects.

• MCP server now exposes update and delete tools for backlog items, so agents can curate the backlog directly.
• Faster task detail terminal: resize and scrollback now load in parallel on mount.
• Activity log is virtualized, so large activity histories render smoothly.
• Board and terminal input pipeline trimmed: per-task lifecycle locks hold for less time, and xterm input is batched into a single IPC write per microtask.
• Backlog dialogs load as sibling components for snappier backlog interactions.

Three new agent adapters

• Cursor CLI - including stream-json init parsing so the agent name and model show up immediately.
• GitHub Copilot CLI - full MCP config integration.
• Warp CLI (Oz) - newly renamed and polished.

Asana board integration

• Import issues and comments from Asana with an OAuth PKCE setup wizard.
• PAT authentication available for attachment-heavy imports.
• Attachments are now captured during import.

Reimagined project sidebar

• New layout with project counts in the header and group rows.
• Per-project idle and active task counts at a glance.
• Polished UX throughout.

Command bar upgrades

• Ctrl+Shift+P now shows the full ContextBar in a responsive overlay, not just the prompt.

Board improvements

• Done column redesign: reclaims worktrees on archive while preserving session resumability, with a clearer move-to-Done confirmation flow.
• Worktree base-branch encoding: auto-generated worktree names include the base branch, and the task UI surfaces it so you can see what each branch forked from.

Agent quality

• Aider: session history parsing, transcript cleanup, and mode-aware idle detection.
• Background shell detection: activity tracking distinguishes user-driven work from background shells to prevent false idle states.

Multi-agent support

• Codex CLI, Gemini CLI, and Aider are now first-class agents alongside Claude Code.
• Each agent has its own permission modes, display name, CLI detection, and session-history telemetry.
• Pick a default agent per project, and override it per column.
• Context handoff between agents on column moves passes the prior agent's native session history to the next agent.
• Session resume works for Codex and Gemini.
• Welcome screen shows every supported agent in the detection grid.

Claude rate-limit quotas

• Task detail context bar shows Claude Code's 5-hour and weekly plan-usage bars.
• Toggleable under Settings -> Terminal -> Context Bar -> Rate Limits.

Changes panel

• Open Changes directly from the task context menu.
• Added to the Command Terminal dialog.
• Auto-selects the first file and shows untracked files.
• Expand/collapse toggle for denser workflows.

Board & UX polish

• Add Column button moved to the board toolbar with a dedicated create dialog.
• Confirmation dialog when moving a task with pending changes to To Do.
• Command Terminal fetches and fast-forward pulls the base branch before spawning.
• New Layout settings tab: card density, column width, panel visibility, window restore, animations.
• Notifications correctly label Command Terminal idle sessions and reopen the overlay on click.

MCP Server

• Rewritten as an in-process HTTP streamable-transport server (replaces the old file bridge).
• New tools: kangentic_get_current_task, kangentic_delete_task, session file/event accessors.
• Unified task creation: routes "create a todo task" to the active board, "Backlog" column to the backlog.
• get_transcript now returns rich structured transcripts.

Session lifecycle

• Session state machine redesigned with atomic transitions.
• Per-task lifecycle locks eliminate pause/resume race conditions.
• Faster task-move to agent-spawn latency.
• Project switching feels instant.

• Git diff viewer in the task detail dialog - see all file changes for any task with split or inline view modes
• Ctrl+V paste support for images in the terminal - clipboard images are saved to a temp file and the path is written to the PTY for Claude Code vision input, with shell-aware path quoting

• File autocomplete in description editors - type @ to search and insert file paths from your project

• Backlog View - stage tasks in a dedicated backlog before promoting them to the board. Includes priorities, labels, drag-to-reorder, filtering, multi-select bulk operations, and right-click context menus
• External Import - pull issues from GitHub Issues, GitHub Projects, and Azure DevOps into the backlog with full metadata (descriptions, labels, attachments, comments)
• Board Filtering - filter board tasks by priority level and label using a new filter popover. A search bar (Ctrl+F) filters across all columns
• Labels and Priorities on Board - labels and priority levels now carry from backlog to board tasks. Edit them in task forms and see label pills on cards
• Markdown Descriptions - task descriptions render full GitHub Flavored Markdown (tables, task lists, strikethrough, links)
• File Attachments - paste or drag-and-drop any file type as an attachment, not just images. Drag files onto the terminal to insert file paths
• Task Display IDs - each task gets a sequential numeric ID (e.g., #42) shown on cards and in the task detail dialog. Click to copy
• Status Bar - aggregate cost and token usage at the bottom of the window with a time period dropdown (Live, Today, This Week, This Month, All Time)
• Context Menus - right-click task cards and backlog rows for quick access to move, edit, delete, and archive actions
• Amber Idle Indicator - idle sessions now show a pulsing amber dot on their tab instead of auto-focusing the panel (auto-focus remains available as an opt-in setting)
• Background Command Terminal - command terminal sessions persist when closed and can be reattached, including across project switches
• MCP Enhancements - display IDs, label color support, file attachment support, and an external command bridge for preview isolation
• Column Drop Highlights - drag-and-drop targets now use the column's configured color for visual feedback

• Ephemeral command terminal - open a quick Claude Code session from the title bar without creating a task
• Automatic PR URL detection - PR links are extracted from terminal output and attached to tasks automatically

• MCP Server for agents - Claude Code agents can now create tasks, search the board, view statistics, and more through structured MCP tool calls during their work. Enabled by default, configurable per-project in Settings > MCP Server.
• Copy Image from attachments - right-click any image attachment thumbnail to copy it to your clipboard
• Sidebar project actions - selected projects now show action buttons (Open, Settings, Delete) directly on the row, plus a right-click context menu with Rename, Open in Explorer, and more

• Commands have been converted to skills with a new skills palette for quick access

• Custom branch names for tasks. Configure the branch name directly when creating or editing tasks, with full branch config UI in the backlog edit dialog.
• Default base branch setting in team-shared kangentic.json, so all team members use the same base branch without per-user configuration.
• Moving tasks out of Backlog is now a destructive move (kills existing session), preventing stale sessions from lingering.

• Project grouping in the sidebar. Organize projects into collapsible groups for better navigation.
• Visual Studio keyboard shortcut preset for Windows users.
• Session duration tracking via heartbeat analytics events.

• Task search bar for filtering tasks across all columns (Ctrl+F)
• Completed Tasks dialog with sortable data table for reviewing finished work
• Redesigned Done column with capped preview and quick access to completed tasks
• Switch base branch or enable worktree on existing tasks (no longer locked at creation)
• "Add task" button now only appears in Backlog for a cleaner board layout
• Settings tab sidebar now shows muted Project/System section headers for clarity
• /pull-request command for creating PRs directly from the board
• npx kangentic --demo flag for trying Kangentic without a real project

• Configurable shortcuts: Add custom command buttons to the task detail dialog header and menu. Launch VS Code, TortoiseGit, terminals, or any tool with one click. Template variables ({{cwd}}, {{branchName}}, etc.) resolve at runtime. Presets for 16+ popular tools included. Team and personal scopes via board config files.
• Claude commands popover: Browse and run project-specific Claude Code commands directly from the task detail dialog. Search by name, navigate with keyboard.
• Session summary for completed tasks: View cost, duration, token usage, tool calls, and lines changed for finished sessions. Sort and search completed tasks in the Done column.
• Task card status indicators: Cards now show contextual status at a glance: model name with context percentage while running, idle icon when waiting for input, "Queued...", "Paused", or "Initializing..." during transitions.
• Mechanical documentation audit: New doc-auditor agent mechanically verifies that docs enumerate all source-code structures (IPC channels, DB columns, config keys, type unions). Integrated into the release, commit, and doc-update workflows.

• Shareable board configuration: Teams can commit kangentic.json to share column layout, colors, icons, actions, and transitions. Personal overrides via kangentic.local.json (auto-gitignored). Live file watching detects teammate changes and offers to reconcile.
• Permission mode guard: Moving a task to a column with a different permission mode automatically suspends and resumes the session with the correct CLI flags. A shimmer overlay indicates the transition.
• Settings search: Filter any setting instantly with a search bar. Multi-token matching, grouped results by tab, match count badges.
• User-paused session persistence: Sessions paused manually via the pause button stay paused across app restarts, respecting user intent.
• Configurable context bar: Toggle individual elements (shell, version, model, cost, tokens, context usage, progress bar) on or off in settings.
• Unified settings panel: VS Code-style layout with Global and Project scope tabs in a single panel.
• Custom Claude Code agents: 5 built-in agents for proactive validation (HMR integrity, IPC audit, migration safety, cross-platform guard, session debugging).

• Tasks created in auto-spawn columns now automatically launch an agent session
• Improved first-launch experience with a welcome overlay and git repository detection
• New Notifications tab in settings with terminal options, idle timeout configuration, and window restore behavior

• Automatic application updates -- Kangentic now checks for and installs updates automatically via electron-updater

• Cross-platform desktop Kanban board purpose-built for Claude Code agents
• Drag-and-drop task management with automatic agent spawning on column transitions
• Live terminal output streaming from Claude Code sessions
• Session persistence -- resume agents across app restarts
• Git worktree integration for isolated agent workspaces
• Hook-based activity log showing structured tool calls and agent state
• Image attachments in task dialogs
• Multi-theme support with 8 named color themes and light/dark/system switching
• Token usage display in task detail and app footer
• Desktop notifications when agents go idle or need input
• Auto-command feature for swimlane columns
• Persistent project reordering via drag-and-drop in sidebar
• Per-project config overrides with global/project settings scope
• Resizable sidebar with auto-collapse
• npx launcher for easy installation (npx kangentic)