← Reference
Keyboard shortcuts
All key bindings, grouped by context.
Full notes from the app
Most shortcuts work only when their pane is focused. Mission Control's app-level shortcuts (⌘N, ⌘K, ⌘1–9, ⌘M, ⌘P, ⌘T, ⌘I, ⌘⇧I, ⌘, and the rest below) are rebindable without a recompile via the top-level [keymap] block in config.toml (see the Configuration section's keymap key); Esc and Ctrl-C stay built-in.
92 entries · kcode 0.6.5097
92 shown
| Keys | Context | What it does |
|---|---|---|
| Global | ||
| Ctrl-L | Global | Toggle the control drawer (open/close). Works even under an open modal, but while a modal (permission prompt, picker) is showing, it owns every other key: a focused drawer pane can't steal the modal's Enter/Esc/arrows. |
| Chat pane | ||
| Ctrl-P | Chat pane | Open the command palette (actions + config + the session's skills, markdown commands and plugin commands, not just slash names). |
| Click an image to enlarge it | Chat pane | Click an image in the chat to enlarge it over the window: an image in a message, or a picture on a tool card such as view_image, a browser capture or a read_pdf page. Another click anywhere, on the picture or beside it, puts it back, and so does Esc. GPU build only. |
| Alt-PageUp / Alt-PageDown | Chat pane | Page the transcript. Plain PageUp/PageDown stay with the composer, along with Home/End, Cmd-Left/Right, and Ctrl-A/E line navigation; chat never captures the input's movement actions. |
| Ctrl-↑ / Ctrl-↓ | Chat pane | Scroll the transcript by 1 line. |
| Ctrl-Home / Ctrl-End | Chat pane | Jump to top / bottom of the transcript. |
| Alt-↑ / Alt-↓ | Chat pane | Focus the previous / next message bubble (the focused one takes a distinct accent focus background; message kinds are told apart by colour, not a gutter bar (your accent-banded message, the model's plain-foreground reply, dim reasoning, code-toned tool rows) see 'Chat text differentiation'). Inert while the palette or a modal is open: the topmost surface owns the keyboard. |
| Ctrl-Shift-C | Chat pane | Copy the active in-app mouse selection if there is one, otherwise the keyboard-focused bubble (Alt-↑/↓), to the clipboard. Writes the native clipboard (the confirmed path) plus best-effort OSC 52 for SSH/tmux; the 'Copied N chars' flash only fires when the native write actually succeeds. |
| Focused chat or note composer | ||
| Composer text editing | Focused chat or note composer | Cmd-A selects the whole multiline draft; typing replaces the selection. Selected text is inverted on a solid theme-accent background with contrasting ink (colors.selected_text or the theme-derived text-on-accent color), rather than a faint wash. Cmd-X/C/V cut, copy and paste; Cmd-Z and Cmd-Shift-Z undo and redo; Left/Right and Shift-Left/Right move or extend selection by whole grapheme clusters, keeping decomposed accents, joined emoji, flags and skin-tone modifiers together. Backspace retains code-point deletion. Other Shift-arrows extend selection, Option-arrows move by word, Cmd-Left/Right and Ctrl-A/E move to line boundaries, Cmd-Up/Down move to draft boundaries. Shift-Enter inserts a newline; Enter submits or files the note. Standard movement belongs to the text input, not transcript scrolling. |
| Chat input | ||
| @file:<path> | Chat input | Attach an image to the turn: write @file:diagram.png (relative paths resolve against the project root, absolute paths as-is) and the image rides on your message. PNG/JPEG/GIF/WebP/BMP/TIFF/ICO are supported; PDFs are not yet. An image wider or taller than 2,000 pixels is shrunk to that size before it is attached (a JPEG stays a JPEG; any other format becomes a PNG); a smaller image is attached as is, except a GIF, BMP, TIFF or ICO, which some provider refuses and is attached as a PNG (a GIF as its first frame). An image whose pixels would take more than 512 MiB once decoded is refused before any memory is set aside for it, and so is a file over 512 MiB, or a path that is not a regular file (a folder, a pipe or a device) before it is read. A model without vision sees a short text note naming the file instead of the image, and an unreadable path surfaces a system notice without blocking the turn. |
| File path in a message | Chat input | Name a text file by its full path in a message (type it, paste it, or drop the file on the session) and your message shows the file's contents under it, as a code block. The contents are read once, as the message is sent, and kept with the session, so the message keeps showing what you sent after the file changes or kcode restarts. The model, exports and search get your message as you wrote it, with the path and not the contents, so the agent reads the file with its own tools when it needs to. Each message shows up to 8 files, 256 KiB a file and 1 MiB in all; a file past those limits, one that isn't UTF-8 text, or a read that takes over 2 seconds leaves the path plain. A path holding a space goes in double quotes, as a drop writes it. A line that starts with a file path is a message, not a command. |
| Mission Control chat input (GPU, maximized) | ||
| Paste image (Cmd/Ctrl-V) | Mission Control chat input (GPU, maximized) | Paste an image from the system clipboard straight onto the turn: Cmd-V / Ctrl-V with an image on the clipboard attaches it instead of pasting text (a text clipboard still pastes text). Both clipboard shapes work: image data (a screenshot or copy-from-browser) and an OS file copy, Cmd-C on one or more image files in Finder, then Cmd-V in the input queues each image file exactly like a drop (non-image files in the copy are ignored, and a copy with no images pastes as usual). The bytes decode off the render frame; pending attachments show as removable chips above the input (one chip per image: 'pasted image' or the dropped file's basename, each with an × that removes just that attachment; the earlier bare count line gave no way to take a mistaken paste back). Same formats and non-vision fallback as @file: (PNG/JPEG/GIF/WebP/BMP/TIFF/ICO; PDFs not yet), and the same shrink: an image over 2,000 pixels on its long edge is shrunk to that size before it is attached. A paste that can't be decoded surfaces a system notice in the panel and is dropped without blocking the turn. A slash command takes the pending images only when it starts agent work from its text: /goal, /loop, /agent, a skill or a markdown command; any other command leaves them in the composer and says so. |
| Drag-and-drop image | Mission Control chat input (GPU, maximized) | Drop image files from the OS onto the maximized session to attach them to the turn; non-image files are handled separately as paths. Each pending image has a removable chip above the input. After Enter, dropped and pasted images appear inline in your message; screenshot names containing spaces or punctuation work too. Click a preview to enlarge it, and click again to put it back. Saved image attachments also preview when you reopen the session, without needing the original file. An image over 2,000 pixels on its long edge is shrunk to that size before it reaches the model. A model without vision sees a short text note instead of the image, and a file that can't be decoded surfaces a system notice in the panel without blocking the turn. |
| Dictation button | Mission Control chat input (GPU, maximized) | Click the microphone at the right of the composer to start the Mac's own Dictation in it, as Edit › Start Dictation does in other apps: the composer takes the keyboard, and macOS listens and types what you say. Dictation is macOS's, so it is turned on in System Settings › Keyboard, and its own shortcut works in the composer too. If macOS doesn't start it, a warning says why. |
| Chat pane (turn streaming) | ||
| Esc | Chat pane (turn streaming) | Cancel the in-flight turn, but only as the base layer: anything dismissible open above the transcript (palette, picker or permission modal, /config editor, focused drawer pane, bubble focus) consumes Esc first, one layer per press. On a permission prompt, Esc denies the tool call and the turn continues with the denial; press Esc again to cancel the turn itself. The mouse has the same one-layer rule: while any overlay is open the grid behind it takes no pointer input (no hover, click, drag or right-click reaches the tiles) and a click on the dimmed backdrop outside the overlay's content dismisses the top overlay exactly as Esc would (a consent dialog such as the project-trust prompt or a provider sign-in stays up until answered; the context menu closes through its animation). |
| Ctrl-C | Chat pane (turn streaming) | Stop the first running thing in the focused session: its ! shell command, else the composer's text, else its turn (as Esc does: a turn parked on a permission or question too). Its subagents are not stopped: they run on, and /jobs stops them. With nothing to stop it quits kcode, and while any session is working, its subagents included, a confirmation names them first; Quit halts their work and the sessions stay resumable. So a ⌃C pressed while a reply streams stops that reply, never the app. Not rebindable. |
| Agent question | ||
| ↑ / ↓ | Agent question | Move the highlight through a question's options and, past the last one, to the Other write-in when the question offers one; the caret moves into the write-in there and back to the composer on an option. |
| Space | Agent question | Tick or untick the highlighted option of a question that takes several answers, while the composer is empty. In a draft, the write-in or a field, Space types a space. |
| Tab / ⇧Tab | Agent question | Move between the options and each of the question's fields in turn; the highlight stays where it is. Back at the options, the caret returns to the composer, or to the write-in while its row is highlighted. A question without fields leaves Tab alone. |
| Enter | Agent question | Send the answer (the highlighted option, or the ticked ones, with the write-in and every field) from the composer, the write-in or any field. |
| Esc | Agent question | Dismiss the question; the agent is told you gave no answer. |
| Chat pane (bubble focused) | ||
| Esc | Chat pane (bubble focused) | Clear focus on the bubble; return to prompt input. |
| Chat pane (default) | ||
| Drag (release) | Chat pane (default) | With mouse capture on (the default), click-drag to select text in-app; releasing auto-copies with a 'Copied N chars' status flash: no Cmd/Ctrl-C needed. A reverse-video highlight shows the selection; clicking never shifts the bubble. Copies message text only: the right-aligned timestamp is a rendering decoration, excluded from both the copy and the highlight. |
| Chat pane (GPU) | ||
| ↓ Latest (jump-to-bottom pill) | Chat pane (GPU) | A floating '↓ Latest' chip appears above the input whenever you are away from the live end, including at the bottom of a loaded page with newer messages still unloaded. Click it (or press Ctrl+End) to release transcript selections, load the newest page and resume following live output. One click lands on the very end, and while you follow, the chat keeps its newest line in view through a window or composer resize, including the composer coming back when a panel closes. Sending a prompt or running a slash command also returns you to the end and resumes following, so the command's reply shows. A fold you open there opens below its header and stops the following when it pushes the newest line out of view. A click during another history load waits for that page instead of being lost. The chip hides only once the newest content is loaded and in view. The history bar's Latest action uses the same recovery path. Only shown on the editable session, never a grid preview miniature. |
| Click a link | Chat pane (GPU) | Hyperlinks and file paths in a message render as clickable, accent-toned, underlined links: click one to open it, a http(s) URL in your default browser, a file path in your editor (the same editor resolution as a /files click, session_panels.files.editor). Both markdown [text](url) links and bare tokens found in prose, bold or italic text included, are linkified, and so is a whole file path written in backticks (`libs/foo.rs`, how the agent usually names one): a backticked path keeps its monospace code-chip look but takes the accent tone + pointer that mark it clickable (no underline: the chip is its affordance), so only a full-span path linkifies and a command like `grep foo src/` stays a plain chip. A http(s) URL links everywhere the chat shows it: inside a code span with other words, in a code block, in thinking, and on every line of a tool card, a command, its description, its output, a diff and a fetched address. A file path in a tool's output stays plain, and so does an address too long for a tool card's header, which the header shows cut short. The file name in a Read / Edit / Write tool-call header uses the same action. File paths, including explicit markdown links, only look clickable when a configured opener or file association is available. Handler discovery runs in the background; unsupported files keep ordinary text or code-chip styling. A click acknowledges the launch with the app name and reports failures, never a false success. Path detection resolves every candidate against the session's project directory and links it only when the file really exists on disk: a relative token like libs/foo/bar.rs or Cargo.toml resolves against the project root, ~/ expands against your home, absolute paths pass through, and the click target is the resolved absolute path, so the editor opens the right file regardless of cwd. A trailing line number rides along: src/main.rs:42 opens at line 42. A slash-path naming an existing directory links too, while a token that resolves to nothing (or prose like e.g. / and/or) stays plain text. URLs linkify by scheme alone, unchecked. A link opens on press (mouse-down on it): the same press-to-act idiom every other transcript control uses, so it fires reliably; a selection dragged in from outside the link still crosses it and selects, but a press that begins on a link opens it rather than starting a selection there. A pure client-side action: opening a browser or editor never reaches the daemon. GPU build only. |
| Click a subagent | Chat pane (GPU) | A click on a subagent's live row, or on the name in its spawn or await card, shows it in the session's read-only subagent tile beside the chat: its saved conversation, live while it runs, under its /agents facts. A session has one subagent tile, so another subagent swaps in. The ✕ in the tile's header closes only the view; /jobs stops the subagent. A row's ✕ removes the row instead, and a grid miniature brings its session forward rather than opening anything. |
| Palette | ||
| ↑ / ↓ | Palette | Move palette selection. |
| Enter | Palette | Accept the highlighted palette entry. |
| Esc | Palette | Close the palette without selecting. Leaves a streaming turn untouched. |
| Slash autocomplete | ||
| Tab | Slash autocomplete | Fill in the highlighted command and a space without running it, the caret after them, so what you type next is its argument. |
| Enter | Slash autocomplete | Run the highlighted command. One that takes text (/goal, a skill with an argument hint) is filled in like Tab instead, and a command typed in full runs as typed. |
| Click a slash command | Slash autocomplete | Does what Enter does for that row. |
| ↑ / ↓ | Slash autocomplete | Move highlight up/down through matches. |
| Prompt input | ||
| Shift-Enter / Alt-Enter / Ctrl-J | Prompt input | Insert a newline; stays in edit mode. The sent message keeps its lines, as a reply does. |
| Enter | Prompt input | Submit the prompt. |
| Esc | Prompt input | Cancel the current input. |
| Prompt input (empty) | ||
| ↑ / ↓ | Prompt input (empty) | Navigate prompt history (persisted per project: survives restarts). While a turn is running, further submitted messages park in the session's client-side queue (dimmed rows above the composer, N queued messages · ↑ edit the newest) instead of being sent; they flush in order as each turn completes, promoting from dimmed rows to real transcript bubbles, and mid-turn, the agent loop pulls the queue at each round boundary: parked lines join the in-flight turn as user messages the model sees on its next round, instead of waiting for the turn to end (attachments ride along, decoded exactly like a fresh turn). The pull is lossless: a parked line leaves the queue and echoes as a bubble only once the daemon confirms it injected that line into the turn, a pull that times out mid round-trip keeps every line parked for the next boundary rather than dropping it. ↑ on an empty composer pops the newest queued message back into the composer as an ordinary draft (attachments included), edit it and Enter resends it as a fresh submission (the back of the queue while still busy, straight to the agent once idle). ↓ on a popped message you haven't changed puts it back at the end of the queue, images included, and empties the composer; with no turn running it stays a draft, because the queue would send it at once. With text in the composer, ↑ walks prompt history as usual and never touches the queue. Prompt history records a queued line only when it actually sends, and a walk in progress stays on the entry it shows. The queue is in-memory and client-side by design: it does not survive closing the window. |
| Settings dir editor | ||
| Enter | Settings dir editor | Submit and save the current field (the Advanced-section Command/Skill dirs text editor). The editor body is a real text field: a caret, cursor keys, selection, ⌘V, ⌘Z; ⇧Enter inserts a newline. |
| Esc | Settings dir editor | Cancel: close without saving. |
| Mission Control (GPU) | ||
| Click a session tile | Mission Control (GPU) | Maximize that session into the live, editable window. Grid tiles are live miniatures: each is the full maximized session view rendered and then uniformly scaled down (one scale factor, so fonts and UI look tiny), a picture-in-picture / Exposé thumbnail, not the view re-laid-out at tile size, which you click to blow up. The click always maximizes immediately. The grid holds only session tiles. When there are more sessions than fit at the smallest tile size, the rows that don't fit run on below and the grid scrolls: the mouse wheel or trackpad over the grid scrolls it, and moving the selection or entering the grid scrolls the selected tile into view. |
| Hover a session tile | Mission Control (GPU) | Resting the pointer on a grid session tile solidifies it: the resting frosted translucency lifts to fully opaque so its content reads crisply, the readability cue that this is the tile you're about to act on. Rendered by Wingman. GPU build only. |
| Empty grid (invitation) | Mission Control (GPU) | When no sessions are on the grid, a calm centred invitation fills the empty grid as a full-window overlay carrying the same deep control-room gradient as a populated grid (so an empty grid reads as the same dark room, not flat slate), naming the next step. The step is whatever isn't done yet: with no model connected it leads with connect a model to begin (a click straight into Settings › Providers, where keys and the ChatGPT sign-in live) with a provider connected but no default model (nor a [[models]] pin) it leads with choose a model to begin, a click into the Default model picker, and once new sessions have a model, press ⌘N to open a project (your own [keymap] chord; the ⌘P palette's New session when it is unbound). Right-click is deliberately not taught here: the context menu still works, it is just not the first thing a new user reads. The prompt softly breathes (a gentle opacity pulse on Wingman's Breathe motion token; static when animations are off: only the prompt breathes, the gradient stays still) and vanishes the instant a session exists. It is gesture-transparent: a right-click or ⌘N passes straight through it to summon the context menu / new-session directory picker; its hitboxes are the connect (or choose) row and the checklist's hide. Under the prompt sits the getting-started checklist: connect a model · open a project · send a message, each ticked (✓) or hollow (○), the first open step in the body colour, a muted hide at the end. Every step is derived from real state, a connected provider with a default model or a [[models]] pin, a session in the registry, a provider round that concluded (the model answered), and never stored, so the row cannot lie: it disappears the moment every step is done, and until the daemon has answered whether a model has ever answered (read off the cost ledger at attach) there is no row rather than one that ticks itself a moment later. Only hide persists (the user store; the daemon replays it on attach). There is no '+' launcher tile: spawning is via ⌘N, the context menu's 'New session', and the palette's 'New session'. |
| What's new | Mission Control (GPU) | The first time a new build runs, the notification center gets one ✦ new row per release-notes bullet you have not seen yet, a bullet added later to a day you already saw included: the text is <date> · <bullet>, plain text, and the drawer reads them as the notes are written, newest section at the top, so a build number that changed since your last launch also says what changed. The notes live in docs/release-notes.md (dated ## YYYY-MM-DD sections, compiled into the binary); the position you have read up to is kept in the user store beside the build version, so the post happens once per build, never per launch, and a build whose notes gained nothing posts nothing. The rows belong to no session: opening the drawer marks them read, they never raise a macOS banner, and the ✕ dismisses them like any other row. |
| License cards (sign-up, code, lock, analytics question) | Mission Control (GPU) | A release build of kcode runs only on a Mac signed up with kcode.sh. Until it is, one card covers the whole window and nothing else can be opened, read or run: Welcome to kcode takes your name and email; Send code emails a code, and Check your email takes the digits, with Send a new code and Use another email. The code card also shows Help improve kcode: what product analytics kcode sends (which features and models you use, how often, and the errors you hit; never your code, prompts, replies, file names or project names), the question Can we include your account?, a What is sent link to kcode.sh, and Include my account and Keep it anonymous side by side, neither picked; Tab reaches them and Enter picks one. The answer goes with the code. Leaving it unpicked is fine: kcode still runs, and the analytics are sent anonymously. The same analytics are sent whatever the answer; the answer decides only whether they are tied to your account. A signed-up Mac that hasn't answered shows the question once in the window: the same question with What is sent and Not now, raised only while no other card or panel is open; Not now, Escape or Enter as it opens closes it on this Mac. Change the answer any time on your account page at kcode.sh. A new account starts its free trial at once: one trial per Mac, so a Mac whose trial another account had says so and offers to buy. When the trial or subscription ends, or the license can't be read, kcode is locked says why and offers Buy a license (kcode.sh signed in as your account), Check again, Use another email and Contact support; while it shows, no session opens and no model call goes out, and the sessions it stopped open again once the license is active. Escape, a click outside the card and the app shortcuts do nothing while it shows. A build made for development needs no license and shows none of this. |
| Single-session grid (centred + hint) | Mission Control (GPU) | When exactly one session is on the grid, its lone tile is centred in the window on its own width: a deliberately-placed little-window rather than a tile stranded in the top-left with an empty half beside it (the grid still floors a lone tile at the two-up 'little window' size so it never goes full-bleed, which would be indistinguishable from the maximized view). A calm muted '⌘N to start another session' hint sits in the dead band below the centred tile, the single-session companion to the empty grid's 'connect a model / press ⌘N to open a project' invitation, and (like it) lives in the passive underlay, breathing softly (static when animations are off) and passing every gesture straight through. Unlike the always-shown empty-grid invitation, the solo hint teaches once and then auto-dismisses after a short linger (session_presenter.solo_hint_linger_ms, default 5000 ms; 0 keeps it permanent), re-showing each time the grid returns to exactly one tile so it is not a fixture over the lone tile. While the getting-started checklist is still open it takes the hint's slot under the lone tile (typically ✓ connect a model · ✓ open a project · ○ send a message) and the ⌘N line's linger starts counting only once the checklist is done or hidden. Both hints track the session count: zero → the invitation, one → this hint, two or more → nothing (a populated grid needs no prompt). The centring is Wingman's generic 'centre a solitary tile' grid option (layout.grid.solo_align), which kcode opts into; the hint content and the count-band policy are kcode's. GPU build only. |
| Attention glow (grid tiles) | Mission Control (GPU) | In the grid, each session tile carries an ambient glow that reads its liveness at a glance, and the glow spills its colour past the tile edge into the surrounding gaps (a soft light-bleed). Routine activity (a session working, or armed to loop) glows the theme accent and breathes on a calm ~4s cycle; every active tile shares one clock, so they rise and fall together as a synced collective, keeping a busy grid calm. A tile that needs you breaks that calm: it switches hue and speeds up to an agitated ~1.6s breath, the theme's waiting colour when it is waiting on your input (a permission prompt or an agent question parked the turn on you: answering it, from any tile, returns the glow to working once nothing else waits on you; a prompt whose subagent you stop from /jobs closes by itself, and the glow goes back to working, or to quiet once the turn is over; a prompt you answer in one window closes in your other windows, and one whose turn another agent stops closes everywhere), the theme's error colour when its last turn failed (until it starts more work: a turn, /compact or an /agent run), so the one tile that wants attention stands out from the synced field. A session whose turn finished while you were elsewhere keeps a steady outline in the theme's success colour, with no breath, until you open it; the first time the grid shows with that tile on it (the kcode window in front and nothing over the grid) the tile shakes once, side to side, for about 0.4 s. Outside the grid, a one-line toast names it instead (Proj B finished), a click on which opens the session, and nothing goes to the notification drawer. Needs input and an error outrank it. Other idle and stopped sessions are quiet: no glow, static and dark. With animations off, active tiles hold a steady lit glow instead of breathing, and a finished tile keeps its outline but does not shake. Rendered by Wingman; the liveness→colour mapping is kcode's. |
| Right-click (context menu) | Mission Control (GPU) | Summon a context-aware menu of actions at the cursor. Right-clicking a session tile lists that session's actions: Close (stops a live session and removes the tile, with a confirmation only while a turn is running; a stopped session also offers Resume, ahead of Close) and Group with… (a picker of the other tiles, the keyboard route into a cluster); a maximized tile leads with Cut / Copy / Paste / Select all for its composer (a click maximizes a tile, Esc / ⌘. restores it, and /compact is a session command and a palette row); a tile in a group also gets the group actions, Focus group (maximizes just that cluster: its tiles fill the window side-by-side, every other tile hidden, and the leftmost holds the keyboard until you click or Tab into another, which then keeps it, Escape returns to the grid), Rename group (a text-editor overlay pre-filled with the current label; the new name persists and re-shows in the cluster's pill), and Ungroup (dissolves the whole cluster). Right-clicking empty grid lists the app actions: New session, Switch, Notifications, Stats, Commands (the palette), Settings, Theme, and the layout modes (Grid / Tabbed / Sidebar). It is a single vertical menu: every action in one column, scrolling when it outgrows the window (no ring, no overflow list). Navigate with ↑/↓ and Enter, or click a row directly; Esc dismisses. Every action maps to the same operation as its command-palette / shortcut equivalent. GPU build only. See #groups |
| Cmd-1 … Cmd-9 (jump to session) | Mission Control (GPU) | Open the Nth session tile in the order the layout shows the sessions (Cmd-1 the first session, Cmd-2 the second, and so on through Cmd-9) skipping any non-session tile. The grid shows each group's tiles together, the groups first, then the ungrouped tiles, and the Sidebar rail and the tab strip list the sessions top to bottom and left to right. Works from the grid (blows that session up into the live window) or from another maximized session (jumps straight to the Nth); in the Sidebar and Tabbed layouts it selects that session's row or tab and the layout stays. A no-op past the last session. A pure client-side view move, never a daemon round-trip, never persisted, exactly like a click-to-maximize. The shortcut lives in kcode: Wingman relays the unbound key. GPU build only. |
| Cmd-M / Cmd+Period | Mission Control (GPU) | Cmd-M opens Mission Control (the grid overview) from a maximized session or from the Sidebar and Tabbed layouts: the discoverable 'way home' chord. Cmd+Period is the Wingman-bound synonym that also un-maximizes back to the grid, regardless of turn state (plain Esc stops a running turn, and itself backs out to the grid only from a maximized session with no turn running; in Sidebar and Tabbed an idle Esc stays). A Cmd+. press never cancels a turn. The layout modes themselves (maximize a tile, tabbed) are switched from the on-screen mode buttons, the settings page, or the context menu, not number chords anymore. GPU build only. |
| Cmd-Shift-] / Cmd-Shift-[ | Mission Control (GPU) | Move the maximized tile to the next (Cmd-Shift-]) or previous (Cmd-Shift-[) session, or a session's /browser tile, in the order the grid shows them (each group's tiles together, the groups first, then the ungrouped tiles) wrapping around the ends, so you move session-to-session without dropping back to the grid. In the grid, where nothing is maximized, they move the keyboard selection instead: the selected tile is ringed with a crisp accent outline so you can see, without the mouse, which tile the keyboard will act on; Return opens it, Cmd-Ctrl-] / Cmd-Ctrl-[ reorder it and Cmd-Return opens its context menu. When the grid has more rows than the window shows, it scrolls to keep the selected tile in view. In the Sidebar and Tabbed layouts they select the next row or tab. Ctrl-Tab and Ctrl-Shift-Tab do nothing. A pure client-side focus change that never touches the agents. Like Cmd-K / Cmd-P the shortcut lives in kcode: Wingman just relays the unbound key. Plain Cmd-[ / Cmd-] without Shift stay with the chat input. |
| Close a session tile | Mission Control (GPU) | Stops that session: its agent ends and the tile leaves the grid (close = stop). Closing a session still working shows a confirmation first (a modal, Cancel highlighted by default); an idle one closes at once. The session stays resumable from the switcher (Cmd-K) or with --resume, except a worktree session whose checkout has no changes or new commits, which is deleted with its checkout and branch while a toast says so (worktree.auto_remove_clean): the same as /exit, the context menu's Close, and the palette's 'Close session'. |
| Drag a tile (reorder or group) | Mission Control (GPU) | Left-drag a session tile: no modifier needed. Pressing a tile lifts it immediately, the card brightens and gains a soft light halo (the Balatro-style pick-up), so it reads as picked up before you move it, and the tile carried under the cursor is a compact lifted card bearing that tile's title. Drop it in a GAP to reorder the grid (the order is saved, so a relaunch keeps it); drop it ONTO another tile to GROUP them into a free-form visual cluster (a soft themed wash + an editable label, see #groups). Drop onto any member of an existing group to add the tile to it; drag a grouped tile out into a gap to remove it. A group keeps its members contiguous through grid reflow. Manage a group (rename / ungroup) from the context menu on any of its tiles. The keyboard twins of reorder and the context menu are Cmd-Ctrl-] / Cmd-Ctrl-[ and Cmd-Return on the tile selected with Cmd-Shift-] / Cmd-Shift-[. GPU build only. See #groups |
| Cmd-K | Mission Control (GPU) | Open the historical sessions: a two-level, project-first overlay for bringing an off-grid session back onto the grid (only off-grid sessions appear; one already on the grid needs no recall). Level 1 lists the projects that have recall-able sessions (the stopped ones and any still running off the grid, across all projects) ranked most-recent-first, each showing its off-grid session count; type to filter by project, Enter drills in. Level 2 lists that project's off-grid sessions under RUNNING / STOPPED headers, the one used last first (a liveness glyph; the session's name: its /rename title, else the first message you wrote in it, else the project directory; and when it was last used): Enter opens the highlighted session, maximized or, in the Sidebar and Tabbed layouts, selected, a running one at once, a stopped one resumed and opened once it is running again, and ← backs out to the projects list (Esc dismisses the whole overlay). A drilled-in project shows its most-recent stopped sessions (session_presenter.switcher.recent_per_project, default 3) behind a 'show older (N)' toggle row; typing a search spans the full history past that cap. When exactly one project has recall-able sessions, level 1 is auto-skipped: ⌘K opens straight into that project's sessions. Closing a tile stops its session, so the usual off-grid session is a stopped one; a RUNNING row appears when something other than you started the session: a fork or a worktree worker spawned from a session not on the grid. With nothing to recall the empty state reads 'No sessions to recall' (not 'No sessions', which read like data loss while a session was live). The switcher's shortcut lives in kcode: Wingman just relays the unbound key. |
| Cmd-P | Mission Control (GPU) | Open the command palette: a flat, type-to-filter list of app / window / session actions, followed by your own skills and markdown commands (grouped under Skills / My commands: choosing one runs the same body as its / slash command, on the focused session), then every slash command this manual lists (grouped under Slash commands, /model, described by the first sentence of its manual entry and then its arguments, Switch the per-session model · [model-id | reset], so a command is findable by what it does: a phrase from a row's description matches the filter too, at half the weight of a name match, and the filter scores the best alignment of the letters you type, tab puts Layout: Tabbed first, because a word start beats the same letters scattered through a longer name or sentence; choosing a command that runs bare (no argument, or optional ones, /version, /model, /worktree) runs it on the focused session exactly like typing it, receipt included, and choosing one whose first argument is required (/rename <title>, /add-dir <path>) puts /rename in the composer ready for the argument). The Help group carries Help (bare /help in the chat), Open user manual (this manual as one page, opened in your browser: written fresh under <data dir>/manual/index.html each time, so it always matches the running build, whose number its footer names), Keyboard shortcuts (⌃B), Version and Quit. Every rebindable app shortcut has a row here: besides New session / Switch session… / Notifications… / Providers… / Check for updates (and, once an update is downloaded, Restart to update) / Layout: Grid (⌘M) / Tabbed, the list carries Jump to your last message (⌘J), Next / Previous session (⌘⇧] / ⌘⇧[), Move tile later / earlier (⌘⌃] / ⌘⌃[), Zoom in / Zoom out / Actual size (⌘+ / ⌘− / ⌘0), Notes… (⌘⇧I, the project's notes overlay) and Note composer (⌘I). Each row shows its live chord, so a [keymap] rebind shows through and an unbound action reads —. The two deliberate omissions are Tile actions (⌘↩ opens the right-click context menu: a menu, whose contents are already palette rows) and the ⌘1–9 jump family (Switch session… is the same operation by name). Like Cmd-K, the shortcut lives in kcode: Wingman just relays the unbound key. |
| Cmd-N | Mission Control (GPU) | Start a new session: opens the in-app folder picker (the same picker the context menu's 'New session' and the palette's 'New session' open), a themed browser in place of the old native macOS folder dialog. Across the top runs the path bar: an ↑ Up button (⌫) that goes to the parent folder, the browsed path with every folder in it clickable (the home prefix collapses to ~), and a + New folder button (⌘N); under it a filter box narrows the folders as you type. The left column starts with Scratch session (a session with no folder until /promote) then lists Places (Home), your Favorites (star any folder with ⌘D or its row's ★: the list persists to session_presenter.favorite_dirs in config.toml, so favorites survive restarts and are hand-editable), and Recents (up to eight projects kcode knows, most-recently-active first; scratch folders, favorites and folders that no longer exist are left out, and two projects with the same name show the parent folders that tell them apart, such as api · work); the right column lists the browsed folder's subfolders, and either column shows a scrollbar while it holds more rows than fit. The footer holds a Hidden folders box (⌘⇧., the Finder idiom; dot-folders are hidden by default) and the Start in <folder> button, which starts the session in the browsed folder: filled and marked ↩ while nothing is highlighted, outlined and marked ⌘↩ otherwise. Mouse: one click on a folder goes into it, a click on a place jumps there, a click on a row's ★ stars or unstars it without going in, and Esc or a click outside the card closes the picker. Keyboard: the picker opens on the Scratch row; ↑/↓ move within a column, Tab switches columns, ↩ goes into the highlighted folder or place (on the Scratch row it starts a scratch session), → goes into the highlighted folder, ← / ⌘↑ / ⌫ go up, ⌘↩ starts in the browsed folder whatever is highlighted, ⌘N makes a new folder (type the name, ↩: it is made inside the browsed folder and the picker goes into it), and ⌘D stars the highlighted folder (on a Favorites row it un-favorites). After going into a folder, jumping to a place, clicking the path or going up, nothing is highlighted, so the next ↩ starts in that folder, even an empty one; ↓ highlights its first subfolder. Typing filters the folders to matching names (a case-insensitive substring) and highlights the first match, so ↩ goes into it; ⌫ edits the filter, the first Esc clears it and only the second closes the picker. A folder that can't be opened (deleted, replaced by a file or locked since the picker listed it) keeps the picker where it is, and a line under the path bar says why (… no longer exists, Can't open …: permission denied); a gone Recents row drops at once. A new folder whose name is taken highlights the existing one and says so, and a name holding / or \, or . / .., is refused with the reason. Adding a plugin from Settings › Plugins opens the same picker titled Add plugin, with no Scratch row and an Add <folder> button. The picker opens in the folder it last browsed for new sessions; adding a plugin from Settings › Plugins remembers its own folder, so it never moves where ⌘N opens. Picking a folder spawns the session there and opens it, ready to type: maximized, or selected in the Sidebar and Tabbed layouts, which stay as they are; the picker feeds the same spawn path the native dialog did, so the project-trust gate applies identically. The native NSOpenPanel remains available: Settings (⌘,) › General › Directory picker cycles App ↔ macOS (session_presenter.dir_picker), live on the next ⌘N. Mission Control holds at most 256 tiles: with every one taken, Cmd-N opens no picker and a warning says to close a tile first. Like Cmd-K / Cmd-P the shortcut lives in kcode: Wingman just relays the unbound key. |
| Cmd-Shift-N | Mission Control (GPU) | Summon the notification center: an edge-anchored drawer sliding in from the right, listing the durable cross-project notification log newest-first (see the /notifications command). The same overlay the command palette's 'Notifications…' and the context menu's Notifications open. A glowing seam on the window's right edge, in the colour of the highest-severity unread notification and with a tab counting them, signals waiting notifications even while the drawer is closed, and a new one shows first as a card in the top-right corner for a few seconds, until the center opens; clicking the seam or its tab opens the center (it is a live affordance, not just an indicator), and visiting a session (focusing its tile) marks its notifications read and dims the seam. Clicking a notification row jumps to the session that raised it (giving it a tile if it is running off the grid, resuming it if it stopped, and opening it: maximized, or selected in the Sidebar and Tabbed layouts) and closes the drawer. Like Cmd-K / Cmd-P the shortcut lives in kcode: Wingman just relays the unbound key; distinct from the bare Cmd-N (new session). Dismiss with Escape or by clicking the dimmed backdrop. GPU build only. |
| Cmd-Shift-S | Mission Control (GPU) | Summon the focused session's /stats dashboard: the overlay with that session's cost/token breakdown, per-turn token + context-occupancy sparklines, tool outcomes and top-tools chart, per-model cost, compaction/continuation/memory counters, and the files the session changed. With no session focused (an empty grid or a non-session tile) it opens the /stats all cross-session dashboard instead: the only sensible stats view there. The same overlay the /stats command and the command palette's 'Session stats' open (the palette's plain 'Stats' is the cross-session dashboard). Like Cmd-K / Cmd-P the shortcut lives in kcode: Wingman just relays the unbound key; distinct from any bare Cmd-S. Dismiss with Escape. GPU build only.See /stats |
| Cmd-T | Mission Control (GPU) | Toggle the focused session's task checklist overlay: one row per task with a bracketed status marker ([x] done, [•] in-progress, [ ] to-do) and its subject, the live view of the task tool's list (see the /tasks command). The same overlay the /tasks command opens; pressing Cmd-T again (or Escape) dismisses it. With no session focused it is a no-op. Like Cmd-K / Cmd-P the shortcut lives in kcode: Wingman just relays the unbound key; distinct from any bare Cmd-T. GPU build only.See /tasks task |
| Cmd-I | Mission Control (GPU) | Toggle the note composer in the maximized session's prompt: the composer switches to note mode, the prompt you had typed is stashed, the rail and a NOTE chip turn the warning hue, and the chip's hint names the live chord, so you can write down an idea for the model later without sending anything now. Enter files the note into the project's .notes/ directory and returns you to the prompt with your stash restored only after saving succeeds; a failure keeps the note draft. Cmd-Alt-I instead files the current prompt directly for later, without entering note mode; Shift-Enter inserts a newline; Escape (or Cmd-I again) backs out keeping the draft, which the next Cmd-I brings back. The slash menu, the @ path completer, and the ↑/↓ history walk are off in note mode, and a note never queues behind a running turn: filing works mid-turn, which is the point. With no session maximized (a grid, or a non-session tile) a toast says so: the note composer lives in the live prompt. Rebindable as note_composer in [keymap]; also in the command palette as 'Note composer'. GPU build only.See /note /notes Cmd-Shift-I |
| Cmd-Shift-I | Mission Control (GPU) | Toggle the focused session's /notes overlay: the project's notes newest first, search by title or body, ↑/↓ to select, Enter to edit. Insert into composer / ⌘Enter explicitly appends the note without submitting; ⌃D deletes after confirmation, ⌃N starts a new note. The same overlay the /notes command opens; pressing Cmd-Shift-I again (or Escape) dismisses it. With no session focused it is a no-op. Rebindable as notes in [keymap]; also in the command palette as 'Notes…'. GPU build only.See /notes Cmd-I |
| Cmd-, | Mission Control (GPU) | Open the fullscreen settings page: the macOS Preferences chord, a synonym for /settings / /config and the context menu's Settings action. An app-level shortcut relayed by Wingman as an unbound key; Esc closes it back to the grid. GPU build only. |
| Ctrl-B | Mission Control (GPU) | Open the keyboard-shortcuts cheatsheet: a static, read-only overlay listing kcode's app-level chords grouped into sections (Sessions, Mission Control, Panels, Zoom, General). It is generated from the same keymap table that drives the real key dispatch, so it always matches the chords that actually work. ⌃B is a free chord (⌃C stops a running turn or quits; every other Ctrl combo is unbound). The same overlay opens from the command palette's 'Keyboard shortcuts'. When it outgrows the window it shows a scroll bar, and ↑/↓/PageUp/PageDown/Home/End scroll it down to its last row. Dismiss with Escape, as its esc Close label says. GPU build only.See #shortcuts |
| Cmd-Q | Mission Control (GPU) | Quit the whole application (the standard macOS Quit chord): a synonym for the command palette's Quit action. While any session is working, a confirmation names the working sessions first; closing the window with its close button does the same. Unlike the other Mission Control chords, which Wingman relays as unbound keys, ⌘Q is wired through the macOS application menu's Quit item: macOS routes the reserved chord through the menu rather than the app's key path (the same reason Cmd+Esc was undeliverable), so the host registers a real app menu carrying Quit and binds ⌘Q to it. GPU build only. |
| Cmd-+ / Cmd-− / Cmd-0 | Mission Control (GPU) | Zoom the whole window: Cmd-+ (or Cmd-=) zooms in by 10%, Cmd-− zooms out, Cmd-0 returns to actual size (50–200% range). Everything grows together (text, panels, spacing, images) the way a browser zooms a page; each change persists session_presenter.ui_scale. The same knob as the Settings page's General → Zoom stepper. App-level shortcuts relayed by Wingman as unbound keys. GPU build only. |
| Mission Control (GPU, chat) | ||
| In-flight tool feedback | Mission Control (GPU, chat) | While a tool is running, its chat row reads as an action in progress rather than a finished one: a present-tense verb (Writing / Editing / Reading / Running / Searching / Fetching / Globbing …) led by a small animated braille spinner in the info tint, a live cue that the step is ongoing, not stalled, so a slow file write or command is never silent. The instant the tool settles the row reverts to its settled verb, which names the action (Write / Edit / Read …), with a far-right ✓ (success) or ✗ (failure) outcome, so a denied write reads "Write … ✗", never that it wrote; a running bash or lsp row shows the spinner where that outcome glyph will land, then swaps it for the ✓/✗ on completion. With animations off the spinner holds its first frame (static): the present-tense verb alone still distinguishes running from done. Gated by /show-tools like every tool row. GPU build only. The pinned working strip above the composer names what the turn waits on: a provider retry ('Provider error — retry 2/4 in 3s', or 'Provider not responding — …' when the provider sent nothing), a park on you ('waiting on permission · bash (general) · 4m' or 'waiting on your answer') relayed by the daemon, live subagents ('2 agents working · web_fetch'), the agent's own set_activity line, or the generic Thinking…/Working… with a '· waiting Ns' stall clock, and Esc interrupts in every one of those states, including a permission park. |
| Chat text differentiation (by colour, not labels) | Mission Control (GPU, chat) | The four kinds of text in the transcript are told apart by colour, weight, and a band rather than author labels: the whole pane stays monospace (it is a terminal), so the distinction is tonal, not typographic. Your own message sits in a subtle accent-tinted band (a left accent border over a soft accent wash) with the message prose itself rendered in the accent colour ('your words in your colour') and no 'You' label. The model's reply is the default foreground with no band, reading as the clean answer. Reasoning is dim behind an indented, muted rail (an aside rather than the reply) and a size smaller: the theme's caption size, scaled with the chat font size, with code inside it shrinking alike. The ✦ Thought header is italic; the body is upright and keeps the model's own bold and italic emphasis. Only the header toggles the fold, so its body can be selected and copied. Its expanded text is rendered Markdown, not flat text, then faded to the theme's emphasis.deemphasized level: code, lists and paths inside a reasoning trace keep their highlighting and structure and simply recede, rather than being flattened to one muted colour. The live trace and the settled fold draw through the same path, so reasoning never reformats itself when a turn settles. Tool rows are the code tone: a monospace header (verb + target) over their monospace output, led by a status-coloured glyph, the accent identity glyph when settled, the info-tinted spinner while running, a green ✓ or red ✗ on the outcome. Every tint resolves from the active theme's accent / foreground / muted roles, so a /theme switch re-colours the whole scheme automatically. Vertical space follows the same logic: a row takes air only where the pair of rows calls for it, so prose turns and open tool output stay airy while runs of one-liners (collapsed folds, finished tool rows, control-plane lines) group tightly into a block instead of each floating on the same fixed gap. GPU build only. The transcript also nests by indent: your message at level 0 with its ▾ turn-control gutter, the assistant's reply one level in (its text lining up under yours), and the reasoning fold a second level in behind a muted left rule; a turn reads as prompt → nested response, not a flush-left wall. |
| Welcome card (first run) | Mission Control (GPU, chat) | A fresh session opens on a centred welcome composition: the floppy mark and tagline above a softly filled, left-aligned prompt; centred ⌘P Commands and ⌃B Shortcuts hints below it. The footer shows the same two hints. Both follow your live [keymap] and omit unbound actions. ⌘K still recalls sessions but is not a primary hint; the redundant /help hint is gone. The tip sits in a quieter centred caption, wrapping inside the prompt's width even for long data-directory paths; short windows scroll the composition rather than clipping its top. In a narrow window the footer takes two rows: the project and model, then the details on one row, leaving out what doesn't fit from the end, the build tag first, then the key hints. It yields to the bottom composer on your first submit, the caret smear snapping from the card down to the bottom chrome. When the presenter reports that no model is connected, the card leads with a connect guide above the input: one muted explainer line plus an accent Connect a model call-to-action that deep-links straight into Settings (⌘,) › Providers, so a brand-new install is one click from connecting a model. Once one is connected the empty composer shows a subdued example (Explain how this repository is laid out) in place of the separate Try action. It remains visible through initial autofocus, disappears when you click the composer or edit it, and stays gone after you clear a draft. It is a placeholder, not draft text: clicking never sends it, and Enter on an empty composer submits nothing. Existing conversations and note mode show no example. After the first turn the welcome card is gone; while nothing is connected a slim No model connected · Connect a model strip sits above the composer with the same click, and a turn sent with nothing connected replies naming Settings › Providers (⌘,) and the ⌘P palette's Providers… row instead of a slash command. The card ends in one tip (muted prose with the thing to type picked out) chosen per session from the tips you have not seen yet: the daemon replays the seen set at attach and until it lands the card shows no tip rather than a repeat; a submit with the card up counts its tip as seen, so over a few sessions the card walks the pool (commands, authored files, config keys, the CLI, integrations) and falls back to the whole rotation once everything is seen. A tip naming a keyboard chord prints your own binding and drops out when the action is unbound; a tip naming a file under the data dir (commands, skills, themes, config.toml) prints it under the directory this launch actually uses: ~/.kcode, or your --data-dir path. Each tip teaches one action in plain language, with an example when arguments matter: /loop 10m check the build repeats a task every ten minutes, and /loop stop ends it. Starting another session and grouping are taught when they apply instead (see Contextual tips). |
| License notice and reminders | Mission Control (GPU, chat) | While the license lets kcode run, each session's status bar can carry a notice: Trial: N days left during the free trial (calendar days on this Mac, reading Trial: ends tomorrow and then Trial: ends today on the last two), Offline: N days left while kcode can't reach kcode.sh to check the license, and Payment failed when a subscription payment didn't go through (the subscription keeps working meanwhile). It turns the warning colour with three days or fewer left, and a click opens your kcode.sh account page. The notification center also gets one ⚠ Account row three days before the trial ends, another the day before it ends, and one per billing period whose payment failed; each is posted once, whatever the launch, and Enter or a click on it opens the account page. |
| Running time (status bar) | Mission Control (GPU, chat) | Once a turn has run, each session's status bar shows how long the agent ran: 4m 12s run · 1h 3m total. run is the time since you last started the agent: a message you sent, a command, /goal, or a /loop you set; a loop's later fires, the goal's nudges, wakes, hooks and timers add to it. total is every turn the session ran. Both leave out the time between turns and the time a turn waited on your answer, count up while a turn runs, and survive a restart. For a session saved before kcode kept them, both are worked out from its turns, without the waits. |
| Contextual tips | Mission Control (GPU, chat) | Six tips are taught at the moment they apply, as one slim row above the composer (the same tip line as the welcome card plus a ✕) or in the welcome card's tip slot while the session is still empty; a submit or the ✕ clears it. On each of your first three app launches, the first session you open teaches /feedback, which sends the kcode team a note; it waits while another tip shows, and the grouping tip waits for it. A window kcode starts again (/reload client, a new build, an update) is not a launch. The first permission prompt teaches /permissions; a turn that ends with the context window past half full teaches /compact; a turn that has been working for a minute (session_presenter.long_turn_tip_ms) teaches ⌘N to start another session while this one works; a second session started from the grid, or from a session maximized from it, teaches in its welcome card dragging one tile onto another to group them: in the Sidebar and Tabbed layouts a drag only reorders, so there the tip waits for a session started from the grid; a turn that concludes in a project with no AGENTS.md teaches /init. Each tip focuses on one action. Grouping is a drag gesture, so its tip still appears when ⌘K is unbound. Each shows only in the session you are in (the new second session excepted, which is the focus), is recorded as seen the moment it shows: in the user store, so it never returns in a later session or launch, except the /feedback tip on your first three launches, and is skipped, not printed with a gap, when the chord it names is unbound in your [keymap]. Nothing shows until the daemon has replayed what you have already seen. |
| Mission Control (GPU, maximized chat or the Sidebar/Tabbed pane) | ||
| Esc | Mission Control (GPU, maximized chat or the Sidebar/Tabbed pane) | Stop the focused session's in-flight turn (Esc=stop-turn, mirroring the terminal UI and opencode/claude-code), but only as the base layer: an open slash menu or overlay (switcher, palette, picker, /control) is dismissed first, one layer per press, and the turn is interrupted only once nothing else is on top. When a maximized session has no turn running (idle, or only its subagents running or asking, which Esc leaves alone) Esc instead backs out to Mission Control: the grid overview (the keyboard twin of ⌘M). In the Sidebar and Tabbed layouts an idle Esc does nothing: the selected pane is where you work there, not a view to back out of, and Cmd-M still goes to the grid. The panel captures Esc in the capture phase so the focused input never swallows it; the presenter then arbitrates (overlay dismiss → turn interrupt → back to grid from a maximized session). Cmd+Period and Cmd-M remain the dedicated 'way home' chords that un-maximize regardless of turn state. A turn parked on a permission prompt or a question is interruptible too: Esc denies the open prompt when one is rendered in this pane, and otherwise sends the interrupt, the daemon denies the permission prompts and questions the turn waits on, so a parked turn never outlives Esc. The turn's subagents run on, even one a spawn_subagent call was waiting for, which reports in a notice when it finishes, and their prompts and questions stay open, as does a question an MCP server asks as itself; /jobs or the chat's jobs card stops them. Esc also stops the session's loops that would start the next turn at once, an on-idle /loop, the /goal driver (the goal is kept) and the agent's on-idle timers, and a timer tick that waited for the turn; a loop on a timer keeps its schedule, and the chat names what stopped and what is still on. |
| Mission Control (GPU, maximized chat) | ||
| Cmd-J (jump to your last message) | Mission Control (GPU, maximized chat) | Scroll the transcript straight to your most recent message (the last instruction) and select it with the same accent highlight a click gives, so you can snap back to what you asked while the agent's reply streams below. A no-op when the conversation has no message from you yet. Not Cmd-↑: the chat input binds Cmd-↑ to 'move cursor to input start', so the jump uses Cmd-J ('jump'). The same jump is also offered as a contextual '↑ Your last message' pill that appears at the top of the transcript whenever that message has scrolled above the viewport, and hides again as you scroll it back into view; clicking the pill does the same jump. The direct-chord twin of the palette's 'Jump to message…' picker, targeting your last message with no overlay. A pure client-side view move, never a daemon round-trip; the chord lives in kcode (Wingman relays the unbound key) and is listed in the Ctrl-B keyboard cheatsheet under General. GPU build only. |
| / then type | Mission Control (GPU, maximized chat) | Typing / at the start of a maximized session's chat input opens kcode's slash-command menu, listing the built-in commands filtered by what you type. The menu opens upward, above the input: kcode's own surface, because gpui-component's completion popover only anchors below the cursor and a bottom-pinned input would clip it. ↑ / ↓ move the highlight (wrapping, via a capture-phase action that preempts the input's own cursor move); Tab completes the highlighted command into the input without running it; Enter runs the highlighted command, only the arg-taking few (/goal, /loop, /rename, /worktree, /add-dir) complete into the input instead, ready for their argument (no more mandatory second Enter). A command typed in full is listed first and highlighted, and Enter runs it as typed, even an arg-taking one or an alias; move the highlight and Enter runs the highlighted row instead. Esc closes the menu. The menu lists the documented built-ins first (aliases folded), then the session's skills, markdown commands and plugin commands, as the daemon sends them (each tagged so it reads as distinct) so they're discoverable rather than typed blind; an entry whose name a built-in or an earlier entry already has is left out (precedence: built-in → skill → command → plugin command). kcode's own commands run in the window; every other /<name> goes to the daemon, which holds the skills and commands and runs it. |
| ~/ then type | Mission Control (GPU, maximized chat) | Typing a token that starts with ~/ in a maximized session's chat input opens a path-completion dropdown above the input, listing the matching files and directories under your home: the shell's ~/… completion, in the composer. The current ~ token (the one under the caret) completes in place; text before and after it is untouched, so a path mid-sentence completes too. ↑ / ↓ move the highlight; Tab or Enter accepts the highlighted path; Esc closes the dropdown. A directory completes with a trailing / and the dropdown stays open so you can drill into it; a file completes and closes. Prefix match on the in-progress segment, case-insensitive, directories listed before files. Dotfiles are hidden unless you type the dot. The menu re-reads the directory on every keystroke, so a path that stops existing mid-type just closes the menu (never an error). GPU build only. |
| @ then type | Mission Control (GPU, maximized chat) | Typing @ in a maximized session's chat input opens a list above the input of the tools the session offers the agent, narrowed by what you type after it. Each row shows a tool's name, and for a plugin's or an MCP server's tool, which one provides it. The @ counts only at the start of the text or after a space or an opening bracket or quote, so an email address opens nothing. ↑ / ↓ move the highlight; Tab puts the highlighted tool's name in with a space after it; Enter does the same, unless you typed the name in full, which sends the message; a click picks a row; Esc closes the list. The @grep stays in your message as written, and kcode adds a <cp:tool-suggestion> note naming the tools the message mentions, so the agent reads them as a suggestion, not an order. The list follows the session's tools as plugins and MCP servers come and go. |
| Drag to select text | Mission Control (GPU, maximized chat) | Click-drag across a message's text to select it character-by-character; the dragged characters highlight (theme accent wash, painted under the glyphs). The drag can span messages and tool blocks: a selection crossing block boundaries highlights the anchor's tail, every whole block in between, and the head block's head, one continuous transcript-wide selection. Tool blocks are selectable too, a diff's code (the line-number gutter and the +/- markers stay chrome, excluded), bash output / command / intent lines, and syntax-highlighted file content, each rendered line a unit, keeping its colours. Inline math contributes its LaTeX source ($…$) and images their alt text, so a span across them copies sensibly. Releasing does NOT auto-copy: press ⌘C to copy (native macOS: the highlight is the feedback, the clipboard only changes on ⌘C). Click elsewhere to deselect. A single click (no drag) selects the whole message instead. |
| Cmd-C | Mission Control (GPU, maximized chat) | Copy to the clipboard, silently (no flash): the active character selection if there is one, otherwise the whole click-selected message's body. A selection within a single block copies the bare slice; a selection spanning blocks copies each block's slice under a ### {header} markdown header, ### user / ### assistant (lowercase role) for a message, ### tool for a tool block, the same convention the session .md export writes, in transcript order, blank line between blocks. With neither selected, ⌘C falls through to the chat input's own copy. Precedence: character selection → selected message → input. Mirrors the terminal UI's copy, native-macOS style. |
| Mission Control (GPU, grid) | ||
| Cmd-Ctrl-] / Cmd-Ctrl-[ | Mission Control (GPU, grid) | Move the selected grid tile (select it with Cmd-Shift-] / Cmd-Shift-[) one place later (Cmd-Ctrl-]) or earlier (Cmd-Ctrl-[) in the order the grid shows: the keyboard twin of dragging a tile into a gap. Inside a group, or among the ungrouped tiles, the tile swaps places with its neighbour. At a group's edge it leaves the group, as dragging it out does, and goes to the front of the ungrouped tiles, which the grid always draws after the groups: it jumps over the next group and never joins it (only dropping a tile onto another joins a group), and the group it left keeps its place. The first ungrouped tile does not move earlier. A no-op at either end (it does not wrap) and in a maximized session (there is no tile to move). In the Sidebar and Tabbed layouts it moves the selected row or tab one place, and a grouped row that lands with no groupmate beside it leaves its group, as a drag in the rail does. The new order and groups are saved, so a relaunch keeps them. Distinct from Cmd-Shift-] / Cmd-Shift-[, which jump the maximized tile session-to-session in Focus mode. A pure client-side reorder that never touches the agents; the shortcut lives in kcode: Wingman just relays the unbound key. GPU build only. |
| Return | Mission Control (GPU, grid) | Open the selected grid tile (select it with Cmd-Shift-] / Cmd-Shift-[), as a click on it does: maximized. It does nothing outside the grid, where the tile on show is already open, or while an overlay is up, which takes Return for itself. Rebindable as open_selected_tile in the [keymap] block; also the command palette's 'Open selected session'. The shortcut lives in kcode: Wingman just relays the unbound key. GPU build only. |
| Cmd-Return | Mission Control (GPU, grid) | Open the context menu (the tile action menu) on the selected grid tile (select it with Cmd-Shift-] / Cmd-Shift-[): the keyboard twin of right-clicking a tile. With no tile selected it opens the app's context menu, as a right-click on empty grid does. A keyboard summon carries no cursor point, so the menu opens in the middle of the window; navigate it with the arrow keys + Enter and Esc to dismiss, exactly like the right-click menu (see 'Right-click (context menu)'). The shortcut lives in kcode: Wingman just relays the unbound key. GPU build only. See #groups |
| Tile identity (window-edge HUD + footer title) | Mission Control (GPU, grid) | Grid cards carry no title chrome. A tile's identity surfaces two ways instead: (1) the window-edge HUD, in the grid, the name of the tile under the pointer (or, with the pointer at rest, the keyboard selection) appears in a short rounded strip centred at the window's bottom edge, sized to the name, drawn from the theme's tokens; at most one strip is ever shown and the tiles themselves carry no label. (2) A session you /rename leads its own chat status bar with the title (title · dir:branch · model · state), and because every grid miniature renders that same footer, a renamed tile is legible in its own card. A session with no model says no model in that status bar and in the sidebar's subtitle, and a click on those words in the status bar opens the /model picker for that session. A session's title is its project directory unless explicitly set with /rename: there is no auto-generated title. Rename via /rename or the switcher's Ctrl+R. GPU build only. |
| Mission Control (GPU, switcher) | ||
| Type, ↑ / ↓ then Enter | Mission Control (GPU, switcher) | Type to fuzzy-filter the current level (case-insensitive subsequence match: project names at level 1, session names at a drilled-in project), ↑ / ↓ to move the highlight through the filtered rows (wrapping), Backspace to edit the query. At the projects level Enter drills into the highlighted project; at a drilled-in project Enter opens the highlighted session (a running one at once, a stopped one resumed and opened once it is running) and ← backs out to the projects list. Esc closes the switcher without recalling. At a drilled-in project, typing also widens the list to the full stopped history past the recent cap. |
| Ctrl+D | Mission Control (GPU, switcher) | At a drilled-in project, permanently delete the highlighted session: its durable row (with the /model, /effort, … overrides), event log, canonical conversation records and both history projections. Raises a confirmation modal first (Cancel highlighted by default); accepting stops the agent if it is live, then erases the session everywhere and removes its tile. Irreversible and NOT resumable: contrast Enter (recall/resume) and closing a tile or /exit (a stop: resumable with --resume, except a clean worktree session, which is deleted, see worktree.auto_remove_clean). Mirrors opencode's ctrl+d delete-session. |
| Ctrl+R | Mission Control (GPU, switcher) | At a drilled-in project, rename the highlighted session in place: opens an edit prompt seeded with the name its row shows, typing edits it (any character, punctuation included), Enter commits the new title (shown in the switcher and the tile chrome), and an empty buffer clears the title, so the row shows your first message again, or the project name. Esc (which dismisses the whole switcher) backs out without committing; only Enter sends a rename. The same daemon-side rename as the /rename command. |
| Ctrl+Shift+D | Mission Control (GPU, switcher) | At a drilled-in project, clear that project's stopped sessions at once: its accumulated pile of finished sessions. Raises a confirmation modal naming the count (e.g. 'Clear 5 stopped'); accepting permanently deletes each stopped session (row + events + buffer). A no-op when the project has nothing stopped. The switcher footer shows the live stopped count next to this key. |
| Mission Control (GPU, palette) | ||
| Type, ↑ / ↓ then Enter | Mission Control (GPU, palette) | Type to filter the palette (case-insensitive substring match on the label), ↑ / ↓ to move the highlight (wrapping), Enter to run the highlighted action. The query is a real text field: a caret, ←/→/Home/End, ⇧-arrows to select, ⌘V to paste, ⌘Z to undo, Backspace to edit. Esc closes the palette without running anything. Commands include New session, Switch session, Notifications, Providers… (provider status, API keys and models: app-level; with a session focused its model-picker tail targets that session, otherwise it opens Settings › Providers), Layout: Grid / Tabbed, Stats (every session's dashboard: Enter on a row opens that session's own) / Session stats (the focused session's dashboard, ⌘⇧S), Toggle thinking / tool calls / timestamps / animations, Switch theme / Switch to light mode / Switch to dark mode / Switch to high contrast, Close / Compact session, Help, Quit. |
| Session composer (GPU) | ||
| Cmd-Alt-I | Session composer (GPU) | File the current composer draft for later as a project note, without submitting or queueing a message. The composer clears only after the save succeeds; a failed save keeps the draft, and text edited while saving is never erased. Works during an agent turn. Notes are text-only: a draft with image attachments is left unchanged with an explanation. In note mode this saves the current note and restores the stashed prompt after success. Rebindable as file_for_later in [keymap]; also available as File draft for later in the command palette (Cmd-P).See /notes Cmd-I Cmd-Shift-I |
| Mission Control (GPU, settings) | ||
| Settings page (sidebar + panes) | Mission Control (GPU, settings) | The fullscreen settings page raised over the grid: a left sidebar of sections and a content pane of live controls for the highlighted section. Click a sidebar row to switch sections, or drive it from the keyboard sidebar-primary: ↑/↓ move the sidebar section, → or Enter dives into its rows, ↑/↓ then move the row and ← steps back out (or nudges a numeric stepper down). Tab is not a section switch here: the GPU window's framework binds it to focus-cycling. A Basic | Advanced switch above the sidebar (⌘1 / ⌘2, or a click) picks which settings the page shows: Basic holds the ones most people change or change often, the personal touches among them (fonts, theme, backdrop, smears), and Advanced holds the rest; each setting shows in one view only. A section with settings in both views ends its Basic list with a link row (14 more Chat settings in Advanced →) and starts its Advanced list with one back (← 2 more Chat settings in Basic); Enter or a click follows it. Account, the Theme designer, Plugins and Providers show only in Basic, Tools only in Advanced. A search covers both views and heads each hit's section with its view (General · Advanced). While a search is up the switch marks the view of the section shown, and ⌘1 / ⌘2 move to that view's first hit. The page opens in the Basic view; ⌘1 / ⌘2 keep you on the same section where it has settings in both views. General (Directory picker, Settings navigation, Motion frame rate and Chat frame rate in its Advanced view): Default model (click to open a searchable picker of every provider's models (the keyed providers first, then the unkeyed ones under Provider · no key sections) led by a “No default” option that clears the default and names what new sessions then start on: the first [[models]] pin, or no model; the row reads (none) while no default is set; the pick persists default_model and relays a live SetDefaultModel to the daemon so new sessions use it without a restart), Default effort (click to cycle Highest supported by model → minimal → low → medium → high → xhigh → max, persisted to default_effort and relayed live the same way), Layout (click the value to cycle Grid→Tabbed→Sidebar; Sidebar is the macOS split view: a left rail lists every session and the focused session fills the rest), Directory picker (click to cycle App ↔ macOS: the in-app browser modal vs the native folder dialog for ⌘N, persisted to session_presenter.dir_picker and live on the next open), Settings navigation (click to cycle Sidebar ↔ Tabs: the page's section nav as the macOS-System-Settings left rail or a horizontal tab bar under the title, persisted to session_presenter.settings_nav and re-laid out in place on the open page), Theme (shows the theme's name; a click opens the theme picker, the same one /theme opens, which lists every theme: kcode-dark, kcode-light, kcode-high-contrast, kcode, marked as following your Mac, and your own; dark mode, light mode and appearance find it in the search), Grid columns (− / + steppers: Auto, the near-square shape that follows the window, then 1–8 tiles a row, with extra rows scrolling; a number you pick holds even when it makes the tiles small), Zoom (− / + steppers, the 50–200% window zoom: same knob as Cmd-+/Cmd-−), UI font and Mono font (each click opens a searchable picker of every font installed on your machine plus a “Theme default” option: the pick persists to session_presenter.ui_font_family / session_presenter.mono_font_family and re-installs the active theme so the new font applies live across chrome and content), Chat font size and Composer font size (− / + in 1 pt steps from 8 to 30 pt; below 8 reads Auto, the theme's size: session_panels.chat.font_size / input_font_size), Animations (on/off: the one switch for all motion), Working animation (click to cycle the eight working styles, dot matrix first: session_panels.chat.working_style), Motion frame rate (click to cycle 15 → 24 → 30 → 60 fps → display rate: session_presenter.animation_fps), Chat frame rate (click to cycle Adaptive → 60 fps → 30 fps: the rate of the chat in the main view, session_presenter.chat_frame_rate), Tile opacity (− / + in 5% steps; 0 reads Auto, the theme's opacity) and Tile blur (− / + in 4 px steps up to 200 px, and down through Off to Auto, the theme's blur), the backdrop rows (Backdrop: click to cycle Shader → Image → Video → Off, the one thing that paints behind the grid, session_presenter.shader.backdrop; Backdrop style, the shader; Backdrop wordmark, the text the dot-matrix shader lights on the empty grid; Backdrop reacts to activity; Per-tile activity light; Background image and Background video, each opening a file dialog whose pick is copied into ~/.kcode/backdrops/ and becomes the backdrop, with a clear action), and the notification triggers (Notify: needs input, Notify: errors, Notify: turn complete, each on/off, relayed live to the daemon). General also holds Editor and Image opener, the commands a /files click opens a file or an image with (session_panels.files.editor / image_editor; unset reads (macOS open) / (Preview)): click a value to open the text-editor overlay (Enter commits and an empty buffer reverts to the default; either way, commit or Esc, the editor peels off and the page resumes at its section (the overlay stack keeps it mounted underneath)). Command files and Skill files (Advanced view): Command dirs and Skill dirs, edited the same way as a comma-separated path list. Memory: Memory auto-recall, a toggle for the per-turn memory.auto_recall gate (off by default): the flip persists to config.toml and relays live through the daemon's generic ApplySettings relay, so running sessions honour it at their next turn. Project hooks (Advanced view): the hooks_enabled switch. Providers: a searchable list led by + Add a provider…, then the providers you added (an Edit badge), then every keyable LLM provider (the six native vendors first, then any OpenAI-compatible vendor), each with a Connected / Add-key badge, type to filter, ↑/↓ to move the highlight, click or Enter a row to open a masked key prompt (a real masked text field: ⌘V to paste, Backspace and the cursor keys edit; Enter saves to ~/.kcode/auth.toml mode 0600 and marks it Connected at once, Esc cancels; either way the prompt peels off and the page resumes at Providers; with no default model set, a saved key or a sign-in then opens the Default model picker over the page at that provider's models). + Add a provider… opens a list: Ollama, LM Studio, vLLM and llama.cpp on this Mac, Another server…, and the hosted services models.dev names that kcode doesn't list (not GitHub Copilot, which takes its own sign-in), and a pick opens a form of Name, Base URL, API and API key filled in from it. The form reads the server's models as it opens and after a change of the base URL, the API or the key, lists them, and warns when a model gets less context than kcode needs (user_providers.min_context_tokens), naming the server's own setting that raises it (or, when the model itself can't take that much, saying to use a larger one), and when a key would travel over plain http. Save checks the server again and saves the provider's [providers.<id>] table to config.toml and its key, if any, to auth.toml; a server that doesn't answer is saved only after Save anyway. Choosing a provider you added opens the same form to edit it, with Remove, which shows a confirmation first. Plugins: the installed-plugin inventory ending in "+ Add plugin from folder…", Enter a plugin for its detail (Enabled toggle, the schema-generated settings form (a secret setting through a masked prompt, kept in auth.toml) source + permissions with the files or hosts each scoped one covers, the tools, commands, skills, agent kinds, output styles and themes it adds to the focused session, the ~1s activity tail, Remove with a confirm for managed folders, none on a built-in); Esc or ← leaves the detail back to the list before a second Esc closes the page. Provider keys live here rather than on a session command because one key serves every session. Layout + grid persist to the daemon workspace; zoom + dirs persist to config.toml and keys and plugin secrets to auth.toml, all off the render frame. Esc from the page closes it back to the grid. The persistent startup default model lives here (all-session: what every new session boots on); the per-session model override and the other session-varying knobs are session slash commands (/theme, /model, /effort, /cot, …), not here. The title bar's search box is a real text field (a caret, ←/→ within the text, ⇧-arrows to select, ⌘V, ⌘Z; it holds the keyboard while the sidebar is active, and ←/→ pressed at its edges navigate: → dives into the highlighted section or jumps to the highlighted hit, and in the Tabs layout ←/→ at the edges switch tabs) and it filters rows, not just sections (the macOS System Settings idiom): typing narrows the sidebar to the matching settings grouped under their section headers (matching on the row's label, its config key, and authored synonyms, in both views) and the content pane previews the highlighted hit's section. ↑/↓ walk the hits; Enter (or a click) jumps straight to that row, the query clears, the section is selected, and the row is highlighted with the keyboard in the content pane. Esc clears the search (then closes). The index is rebuilt from the panels' live rows on each keystroke, so async-filled sections (Plugins, Providers) are findable the moment their rows land. GPU build only. |
| Trust dialog | ||
| ← / → / Tab / ⇧Tab | Trust dialog | Move the highlighted choice on the project-trust dialog (wrapping; Tab steps forward, ⇧Tab back). Opens highlighting This session only (the safe default) or Not now on a question about code a running session holds back. The highlighted choice is the filled button (the one Enter presses) and the others are outlines. The provider sign-in dialog takes the same keys. See the Permission Modes section's Project trust entry for what the dialog gates. |
| ↑ / ↓ / Page Up / Page Down / Home / End | Trust dialog | Scroll the dialog's list of what the project holds. The list grows with the window; when it holds more than fits, a scroll bar shows beside it, a line under it counts the items out of view (3 more below), and the key hint adds ↑/↓ scroll. The provider sign-in dialog scrolls its details the same way. |
| Enter | Trust dialog | Choose the highlighted button (Trust & remember / This session only / Open untrusted, or Not now for code a running session holds back). |
| Esc | Trust dialog | On a spawn/resume gate: cancel the action outright (invoking it again shows the dialog again). On the boot prompt: open this window untrusted, the project's .kcode content stays ignored. On code a running session holds back: Not now. |