* ui: add a voice engine for capture, peaks and transcription Voice input needs a live waveform, a clip that can be sent and drawn, and a transcript, whether the model streams text while the user speaks or not. This adds the engine for it in src/lib/voice, with no imports of app stores or components, so it can move into a package later. - VoiceCapture: one microphone start gives live levels (RMS on a dB curve, fixed history, newest last), PCM16 mono chunks at 16 kHz from an inline AudioWorklet (ScriptProcessor fallback), and on stop a 16-bit mono WAV clip with its duration and peaks. Tracks, nodes and the context are released on stop, cancel, a failed start and page hide. - computePeaks / computePeaksFromAudio: normalised peaks of samples, a buffer, or a stored base64 or Blob file. - transcribeClip: POST /v1/audio/transcriptions, with the server's own error message on failure. - RealtimeTranscriber: the /v1/realtime?intent=transcription socket, queueing audio until session.created, a final commit, and a typed RealtimeUnsupportedError when the model cannot stream or the socket is refused (a browser websocket cannot send the API key header). - startVoiceSession and VoiceRecorder: try realtime, fall back to one HTTP transcription of the clip, and expose levels and live text. Assisted-by: Claude (Opus 5.5 Medium) * ui: add voice waveform and transcript components to Llama UI Voice input needs three visuals that the app can feed with its own recording and transcription logic: the live waveform while recording, a playable waveform of a recorded clip, and a clip with its transcript that opens on hover, keyboard focus or a tap. VoiceWaveformLive draws input levels as many bars as fit the width, newest on the right, and holds still with reduced motion. VoiceWaveform plays a clip through its own audio element, fills the played part in the primary colour, seeks by press, drag and arrow keys (a slider) and pauses any other clip that was playing. VoiceTranscriptReveal keeps the transcript in the DOM for screen readers while it is closed, and shows transcribing and error states. The waveform math (resampling peaks, bar capacity, seek and time formatting) is in src/lib/voice with unit tests. The components are exported from the package root and from @ggml-org/llama-ui/voice, which Llama Chat resolves to the sources like the other entry points. Assisted-by: Claude (Opus 5.5 Medium) * ui: add stories and tests for the voice components Storybook gets UI/Blocks/Voice: the live waveform fed by a fake microphone meter in a chat form row, the clip waveform in its sizes and states, and the transcript reveal in each state and on touch. The demo audio is a WAV synthesised in the story, so no binary file is added. The play functions check seeking by keys and by press, the reveal on hover, keyboard focus and a tap, and a clip without audio. Pressing play needs a real click before a browser plays audio, so a client test covers play and pause and one clip pausing another. Assisted-by: Claude (Opus 5.5 Medium) * ui: store voice messages with their transcript and send it to models that cannot hear A voice message recorded in the chat form is an audio attachment like an uploaded file, plus what the app needs to draw and read it. The audio attachment gets four optional fields: voice (recorded in the app), transcript, peaks and durationMs. Stored rows, export and import carry them as they are, so no migration is needed and older messages keep working. The transcript is for reading. A model with audio input still gets the audio alone; a model without it (the user switched models later) gets the transcript as a labelled text part in place of the audio. When the model's modalities are not known yet, or there is no transcript, the audio goes as before. A transcript often comes back after the message was sent. The new voiceTranscriptsStore follows each pending transcript by the clip's name, hands it to a send that happens after it arrived, and otherwise writes it into the stored attachment of the message it went out with, in the open chat or, if the user left it, in the database. Copying a message now includes its voice transcript. Assisted-by: Claude (Opus 5.5 Medium) * ui: add Voice settings and resolve what the microphone does per model The user should not have to think about which kind of voice input is in use, so the app decides it from the conversation's model and two settings. resolveVoiceMode is a small pure function: with Voice input on Automatic and a chat model that has audio input, a recording is a voice message, transcribed by that same model so a router swaps nothing. Otherwise it is dictation, with the chat model when it can hear, else, in router mode, with the chosen transcription model. When no model can transcribe, the microphone is unavailable and says why. Settings get a Voice section of their own rather than more rows in General, which is already long: these four belong together and one of them lists the server's models. - Voice input: Automatic (default) or Always dictate. - Transcription model: Same as chat model (default) or, in router mode, any model with audio input; on a single-model server it only says whether that model can transcribe. It notes that dictating with another model loads it beside the chat model. - Transcribe voice messages: on by default. - Transcription language: empty for automatic. The keys live in the settings registry like the others, so defaults, storage, reset and export need nothing new. The model picker is drawn by SettingsChatVoiceTab, the way Extensions and Tools draw their sections. Assisted-by: Claude (Opus 5.5 Medium) Co-authored-by: Nicolas Mowen <14866235+NickM-27@users.noreply.github.com> * ui: record voice messages and dictation from the chat form The microphone now does what the conversation's model allows, without the user choosing a mode: its label says "Send a voice message" or "Dictate", and when no model can take audio it is disabled and says why (shown in router mode only when some model there has audio input, so the reason can be acted on; a single-model server without audio keeps the send button as before). The microphone errors keep their warning state and toast, now also for the voice engine's own capture errors. Recording goes through the voice engine's VoiceRecorder instead of the MediaRecorder flow. While it records, the actions row becomes a recording row: cancel, the live waveform, the elapsed time and stop. Escape or cancel drops the recording. The form keeps its two rows while recording, so the message stays visible on desktop and the row fits the phone layouts and the expanded sheet. - Dictation writes the live transcript into the message at the caret as it streams. DictationSpan replaces only the dictated stretch, moves it when the user types before it, keeps what is typed after it, and hands text the user edits over to them. When nothing streams, the row shows Transcribing... after stop and the text goes in once it is back. Cancel takes the dictated text out again. No audio is attached. - A voice message streams its live text, if any, under the waveform, not into the message. On stop the clip is attached as a voice file with its peaks and length, and its transcript is followed by the voice transcripts store (unless Transcribe voice messages is off). Realtime transcription is skipped for servers reached through a tunnel or with an API key, which a browser websocket cannot carry. The first dictation with a separate router model shows a one-time note that it loads beside the chat model. The form takes an optional voiceRecorder, so tests and stories can run it without a microphone. Assisted-by: Claude (Opus 5.5 Medium) * ui: draw audio attachments as playable waveforms with their transcript A voice message should read as a clip, not as a file chip. Audio attachments now render with Llama UI's VoiceWaveform inside VoiceTranscriptReveal, through a direct branch in the core attachment item (voice is core, so it does not go through the extension renderers): the removable sm chip in the chat form, and the md clip in a sent message, on the user bubble's surface. - The audio plays from an object URL made from the stored base64 (or the uploaded file's data URL), revoked when the clip leaves the screen. - A clip stored without peaks, an older voice message or an uploaded audio file, gets them computed from its audio once per session, in memory only; uploaded files keep their name under the waveform. - The transcript opens under the clip on hover, keyboard focus or a tap, and shows Transcribing... while the voice transcripts store waits for it, Transcription unavailable when it failed, and nothing when none was asked for. - The full-screen preview keeps its player and shows the transcript. Tests run the chat form with a fake recorder: the microphone per mode, the recording row, dictation around typed text, the Transcribing... state, cancel, and a voice message chip whose transcript arrives later and goes with the send. Stories show the form recording each kind, a recorded chip, and sent voice messages with the transcript, pending, and computed peaks. Assisted-by: Claude (Opus 5.5 Medium) * ui: drop the old audio recorder and proxy realtime transcription in dev The voice engine replaced AudioRecorder, so it goes. The dev server now proxies WebSockets on /v1, so realtime transcription works under vite dev. The VoiceCapture test waits for the tone to be recorded instead of a fixed 700 ms, which a loaded test run could miss. Assisted-by: Claude (Opus 5.5 Medium) * ui: dictate with a model that is already loaded when the chat model cannot hear The Transcription model setting's Automatic choice now falls back to a router model with audio input that is already loaded, so dictating with a text-only chat model works without picking a model and loads nothing. A picked model still wins, and the empty value is a named constant. Idea from upstream PR 27567, by NickM-27. Assisted-by: Claude (Opus 5.5 Medium) Co-authored-by: Nicolas Mowen <14866235+NickM-27@users.noreply.github.com> * ui: fill the recording waveform, and open voice message transcripts - The live waveform fills the row from the start: a recording shorter than the row is stretched over all its bars, a longer one scrolls with the newest level on the right. No dotted placeholder. - A voice message's transcript is open under its clip, in the chat form while it comes in and in the sent message, with the toggle beside it to fold it. - Settings > Voice > Collapse transcripts (off by default) keeps them folded instead, opening on hover, keyboard focus or a tap. Assisted-by: Claude (Opus 5.5 Medium) (cherry picked from commit 2a93f2f948ed976e40f693ce5d3acf1168284b05) * ui: record voice without ducking other audio, from the first word - Voice input asks for the raw microphone signal: no echo cancellation, noise suppression or auto gain. Echo cancellation made the system treat a recording as a call, which lowers or interrupts other playback, and auto gain faded the first words in. - One AudioContext serves every recording, with the capture worklet loaded once, suspended in between and, where the browser allows it, without an output device, so recording never touches the speakers. - The chat form prepares it when it gets focus (no microphone, so no recording indicator), and a mouse starts recording on press. - A recording starts counting once audio flows: until then the bar says Starting... instead of running the timer. Development builds log how long each step of a start took. - Settings > Voice > Microphone: Automatic (the system default, or the built-in microphone when the default is a Bluetooth headset, which would switch to its low-quality call mode), System default, or a listed microphone. Assisted-by: Claude (Opus 5.5 Medium) (cherry picked from commit 76643e7210d21617d734fdfde677e84af7273153) * ui: show voice text live in the message box, and follow transcripts as spoken - A model that hears but has no live ASR (e.g. LFM3-Omni) now gives text while you record: the realtime client commits the audio piece by piece, at the pauses the levels show, one piece at a time, and the final commit takes only what is left. - A voice message's live text shows as the empty message box's placeholder. Clicking into the box makes it text to edit there, and what is said next follows it. On stop the text leaves the box and goes with the clip; when it was edited, the edit is the clip's transcript. - Llama UI gains VoiceSpokenText and spokenLength(): spoken words in the full foreground colour, which they keep, the rest muted. A clip's transcript follows its playback (estimated over the parts with speech, as there are no word timings), and live text brightens as it arrives. - The recording bar has a send button that stops and sends in one go. Assisted-by: Claude (Opus 5.5 Medium) (cherry picked from commit ce773017f902f8a5690ab9ecca368cecb6427422) * ui: let the capture test start with silence A capture starts with its first audio block, which can come before a stream's first sound, so the steady tone is checked from where it begins. Assisted-by: Claude (Opus 5.5 Medium) (cherry picked from commit a41ee5337b83fef685a5cbaff176925f1647ce9d) * ui: show live voice text sooner, and keep silence from becoming "yeah" - A model that transcribes per commit (LFM3-Omni) gets shorter pieces: cut at a 0.35 s pause after speech, or at the next quiet moment once a piece reaches 3 s, so the text follows the speaker within a few seconds in dictation, the message box and voice mode. - When the recording stops, the audio after the last piece is committed only if speech was heard in it. LFM3-Omni transcribes a commit of pure silence as "Yeah." (or "はい。"), which ended up at the end of the text. A recording whose only piece is that rest still commits it. Assisted-by: Claude (Opus 5.5 Medium) (cherry picked from commit 43e52cbab0fba9d8e84b9491890a406e3c6c90f3) * ui: transcribe a recording cut at pauses once more as a whole Live text comes piece by piece, cut at the speaker's short pauses, and a model reads each piece as a sentence of its own: "I went. To the store." When the recording ends, one pass over the whole clip gives the final text, punctuated by meaning, in dictation and in voice mode alike. The pieces' text stays the live text until then, and is kept when that pass fails or the recording is longer than two minutes. Assisted-by: Claude (Opus 5.5 Medium) (cherry picked from commit 359dd9632e6c88000a1d23c2575bd1e6fcdd9ec1) * ui: give live text over HTTP for models without a realtime session A model that hears but cannot hold a realtime session, such as Qwen3-ASR, was transcribed only after the recording, so dictation showed no text while speaking. The recording is now cut at the speaker's pauses as with a realtime session, and each piece sent to /v1/audio/transcriptions on its own, one at a time, its text joined live. The audio is kept from the start, so what is said while the socket answers goes with the first piece. The whole clip is still transcribed once more at the end, and the pieces' text is kept if that fails. A server a websocket cannot reach (an API key, a remote server) gets live text the same way. Assisted-by: Claude (Opus 5.5 Medium) * ui: say when the browser processes the microphone despite the raw request The browser keeps one processing setting per microphone: while a call, another tab or voice mode has it open with echo cancellation, a dictation that asked for the raw signal is processed too, sounds distant, and the system may lower other sound. The capture now warns when it gets processing it did not ask for, and its development log names the microphone, its rate and what processing is on. Assisted-by: Claude (Opus 5.5 Medium) * ui: keep the microphone beside the stop and send buttons While a reply streams, or once the message has text, the microphone no longer gives way to the main button: it stays beside it as a quieter, secondary button, so more can be dictated without clearing the message or waiting for the reply. Assisted-by: Claude (Opus 5.5 Medium) * ui: show the voice components' stories with Llama UI VoiceWaveform and VoiceTranscriptReveal are Llama UI components, so their stories move to the library's Storybook, under Llama UI/Voice, with the demo audio they play. Llama Chat's Storybook shows them there too, and its own voice stories take the demo audio from the library. VoiceWaveformLive stays with Llama Chat, as its demo records through the chat form's microphone. Assisted-by: Claude (Opus 5.5 Medium) * ui: add a noise suppression setting for voice input Recordings stay raw by default. Settings > Voice can now switch the browser's noise suppression on; echo cancellation and automatic gain stay off either way, as they interrupt other sound and fade the first words in. Assisted-by: Claude (Opus 5.5 Medium) * ui: keep live voice text coming when the room is never quiet Pieces were cut only at a quiet moment, so with music or talk in the room none was ever cut and no text showed until the recording ended. A piece half as long again as its longest length is now cut where it is. Assisted-by: Claude (Opus 5.5 Medium) * ui: add echo cancellation and automatic gain settings for voice input Settings > Voice now has a switch for each kind of microphone processing the browser offers. All three stay off by default, so recordings are raw; each one says what switching it on costs. Assisted-by: Claude (Opus 5.5 Medium) * ui: put the microphone after the stop button, filled like it While a reply streams, the microphone is the last button of the row and looks like the stop and skip buttons beside it, instead of a bare icon before them. Assisted-by: Claude (Opus 5.5 Medium) * ui: say when dictated text is being refined, and send it refined After a recording with live text, the whole clip is transcribed once more so pauses do not end sentences. The form now says so while that pass is out, and a send asked for meanwhile waits for it, so the message goes out with the refined text rather than the pieces. Assisted-by: Claude (Opus 5.5 Medium) * ui: keep the send button in place and put the microphone beside it The send button no longer gives way to the microphone on an empty message. The microphone stands before it as a quieter, tertiary button, with the same tooltip, warning and recording states as before. Assisted-by: Claude (Opus 5.5 Medium) * ui: let a recording be sent while a reply streams, once it has words While a reply streams, the recording row now shows its send button as soon as the first transcribed words are in the message, so the recording no longer has to be stopped first. The microphone also stays beside the send button once a voice message is attached. Assisted-by: Claude (Opus 5.5 Medium) * ui: raise quiet speech for transcription, and say when it is too quiet A quiet microphone gave the model near silence to transcribe. A new Voice setting, Boost quiet speech (on by default), raises speech to a level speech models hear well, up to eight times louder: the audio streamed for live text as it goes, and the whole recording in one pass when it stops, so its first words get the same gain as the rest. Only speech is measured, never the silence around it, loud speech is left as it is, and a soft limiter keeps raised peaks from clipping. When speech stays quiet even at the highest gain, the recording row says so. Assisted-by: Claude (Opus 5.5 Medium) * ui: keep the voice engine free of app code The engine in lib/voice imported the microphone messages from the app's utils and the WAV type from its enums, the only two things keeping it from moving into a package of its own. The microphone checks and messages move into the engine (voice/support.ts), next to the capture that needs them; the uploaded-file helpers stay in the app's utils. Assisted-by: Claude (Opus 5.5 Medium) * ui: start voice sessions from one voice service How a recording reaches the server (routes, API key, whether the realtime socket can be used) and how the Voice settings become capture and transcribe options were built inside the chat form's voice hook, and copied into every other feature that records. VoiceService builds them in one place: a feature asks it for its session options and adds only its own callbacks. Assisted-by: Claude (Opus 5.5 Medium) * ui: tell speech from the room in one place The pause detector, the speech gain and (in voice mode) the listener for speech over a reply each followed a noise floor with their own numbers. A SpeechDetector now does it for all of them, in dB, with one set of settings in VOICE_SPEECH. Assisted-by: Claude (Opus 5.5 Medium) * ui: let any text field take dictation Writing dictated text into a field (at the caret, around the user's typing, replaced by the whole recording's text when it comes) moves out of the chat form's hook into a plain createDictation(). The hook keeps what is the chat form's own: voice messages, toasts and the hint. Assisted-by: Claude (Opus 5.5 Medium) * ui: move the recording row into Llama UI VoiceRecordingBar draws the row shown while recording (cancel, waveform and time, send, and a caption or warning under it) from data alone, so voice mode or a notes page can show the same row. The chat form passes its labels and its microphone button as the stop button. Assisted-by: Claude (Opus 5.5 Medium) * ui: keep the voice engine's Svelte recorder behind its own entry VoiceRecorder, the one part of lib/voice written with runes, moves to lib/voice/svelte, so the rest of the engine is plain TypeScript any app can use without Svelte. Assisted-by: Claude (Opus 5.5 Medium) * ui: move voice input into its own package, Llama Voice The voice engine (capture, the speech signal, transcription over the realtime socket or HTTP) and dictation into a text field move from the app to ui/lib/voice as @ggml-org/llama-voice, so another app, such as a notes app, can use them. The Svelte recorder is its ./svelte entry. Llama Chat compiles the sources, as it does Llama UI's, and keeps its server and settings wiring. The package has its own type check, tests and build, and a CI job for them. Assisted-by: Claude (Opus 5.5 Medium) * ui: show which model does what in voice input The model picker and the Models manager mark each model that hears: Dictation on the one that transcribes dictation, Speech to text on transcription models such as Qwen3-ASR, and Voice on chat models that take voice messages. A loaded model kept only for dictation sits apart from the chat models, a model can be made the dictation model from its row, and Settings > Voice groups the transcription models apart from the chat models with audio input. A voice store now resolves what the microphone does, for the chat form and the model lists alike. Assisted-by: Claude (Opus 5.5 Medium) * ui: warn of a quiet microphone inside the recording row The too-quiet line under the waveform made the form grow and jump. The warning now stays inside the row: the waveform turns the warning colour, an icon sits by the time, and the advice opens above the row in a tooltip for a few seconds, then on hover or focus. Screen readers still hear it when it appears. Assisted-by: Claude (Opus 5.5 Medium) * ui: open the microphone while the pointer is on its button The first seconds of a recording could come out quiet while the microphone settled, and the too-quiet warning showed for them. With the pointer or the focus on the microphone button, the microphone now opens ahead of the press (only when the page may already use it, so nothing asks), and the recording takes it at once; it is let go shortly after the pointer leaves without a press. The warning also waits until the speech has stayed too quiet for two seconds. Assisted-by: Claude (Opus 5.5 Medium) * ui: tidy voice input after review Fix a microphone left open when a recording is cancelled while it starts or the pointer leaves while it opens, a capture that is still letting go silencing the next one, and live text stopping after a realtime error. Share the dictation model choice, the transcript lookup, base64 decoding, PCM joining and stopping a stream instead of repeating them, and export only Llama Voice's public API. Assisted-by: Claude (Opus 5.5 Medium) * ui: pin Llama Voice's dev dependencies to Llama UI's versions Assisted-by: Claude (Opus 5.5 Medium) * ui: let demos speak into a synthetic microphone Assisted-by: Claude (Opus 5.5 Medium) * ui: keep the microphone for voice input, off the audio files entry The attach menu's Audio Files entry used a microphone, which now sits beside the form's own microphone button. The microphone stands for recording and dictation only; a model's audio modality is headphones, and the audio files entry is a file with headphones. Assisted-by: Claude (Opus 5.5 Medium) * ui: set voice input on the slimmed design tokens Voice input's components use the type roles, the one corner scale and shadcn's utility names for the neutral colours, as the rest of Llama UI and Llama Chat do since the design tokens were slimmed down. Assisted-by: Claude (Opus 5.5 Medium) * ui: name voice states, platform strings, patterns and tunings Addresses the review on the voice PR: - Llama Voice: VoiceCaptureState, VoiceRecorderStatus, RealtimeState and RealtimeUnsupportedReason are enums (src/enums.ts, exported); DOMException names, AudioContext states, transcription form fields, socket protocols, device kinds, worklet messages, error messages, regexes and tuning values are named constants. - Llama UI: VoiceTranscriptStatus, VoiceWaveformSize and VoiceRecordingBarStatus are enums; KeyboardKey, MouseButton and PointerType join its UI enums; the voice components' sizes, labels, timings, selectors and patterns move to constants/voice.constants.ts. The voice unit tests move from src to tests/unit. - App: ChatFormVoiceState is an enum, the recording bar's status comes from a map, and the voice code uses the enums above, BooleanString, HEADERS and new voice constants in place of inline strings, regexes and numbers. Assisted-by: Claude (Opus 5.5 Medium) * ui: keep all of Llama UI's tests under tests The Model UI formatting spec moves to tests/unit and the button spec to tests/client, next to the voice specs; the test projects now read only tests/unit and tests/client. Assisted-by: Claude (Opus 5.5 Medium) * ui: show a transcript's words without the model's own format Qwen3-ASR names the language before its text (language French<asr_text>...) and VibeVoice-ASR answers with timed segments as JSON, which the server passes on as they are. The voice engine now takes both off, over HTTP and over the realtime socket, also while segments still stream, so the message box and the clip show the words alone. Assisted-by: Claude (Opus 5.5 Medium) * ui: dictate when a speech-to-text model is picked for the conversation A speech-to-text model has audio input, so on Automatic the microphone recorded a voice message for it, which such a model has nothing to answer with. It now dictates, as the model's badge already says it is for. Assisted-by: Claude (Opus 5.5 Medium) * ui: say why a live-only model cannot transcribe behind an API key A browser websocket cannot send the API key, so the realtime socket is left out on a server that needs one, and a model that only transcribes over it (Voxtral Realtime) answers the HTTP route with a 501. The clip and the dictation toast now say so instead of only that transcription is unavailable. Assisted-by: Claude (Opus 5.5 Medium) * ui: pick the dictation model with a microphone on its row The model that transcribes dictation was marked with a filled Dictation badge. It is now marked by a red microphone among the row's actions in the model picker, beside the heart, and every other model that hears shows the same microphone on hover: a click makes it the transcription model, and a click on the chosen one hands the choice back to Automatic. The badge keeps saying what kind of hearing model a row is, and the Models manager marks the dictation model with a small microphone beside its name. Assisted-by: Claude (Opus 5.5 Medium) --------- Co-authored-by: Nicolas Mowen <14866235+NickM-27@users.noreply.github.com>
61 KiB
version, name, description, colors, typography, rounded, spacing, components
| version | name | description | colors | typography | rounded | spacing | components | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| alpha | Llama UI - Matte Glass | Llama UI, the design system and components of Llama Chat, the chat client llama-server bundles. Neutral and quiet: frosted matte surfaces with faint blue-grey hairlines, near-flat fills, and status colours for success, warning, error and info. Tokens come in three tiers: primitives, semantic tokens (with a "-dark" twin where dark mode differs and a "-matte" twin where the Matte light theme does) and components. The colors, typography and rounded sections are generated from ui/lib/ui/src/lib/styles by `npm run design:tokens` in ui/apps/chat; spacing, components and the prose are kept in sync with the CSS by hand. |
|
|
|
|
|
In short
The rules most changes need. The sections below hold the detail and the reasons, and they win where this list is too short. Before adding a token, a size or a component, use what is here; if something really is missing, add it here first.
- Components: take base components from Llama UI (
@ggml-org/llama-ui/ui/<name>); do not build a new one when an existing one fits. Style them through their variants (variant="outline",size="sm"), not by overriding classes. - Colour: greys build the UI; colour is for meaning only (status, destructive actions,
focus). Neutral colours go by shadcn's utility names (
bg-background,bg-muted,bg-card,text-foreground,text-muted-foreground,border-border), status bybg-surface-*,bg-fill-*,text-fg-*andborder-line-*-softofsuccess,warning,dangerandinfo, and any other semantic token asbg-(--surface-lifted). In CSS, read the semantic token (var(--surface-page)). No Tailwind palette classes, no hex. - Type: Llama Sans, and Llama Mono for code, through roles:
text-body(16px) for chat messages and the chat input,text-body-sm(14px) for most text,text-label(14px, medium) for buttons and labels,text-label-smandtext-caption(12px) for badges and hints,text-caption-sm(10px) for tags and counts,text-title(18px) andtext-heading(24px) for the rare heading,text-display(30px) for the biggest headline. Never Tailwind's own sizes (text-sm,text-xs) or a size in brackets. - Shape: one corner scale,
rounded-smtorounded-4xl:rounded-md(8px) for controls, fields and badges,rounded-lg(10px) for menus and dialogs,rounded-xl(14px) for cards,rounded-2xl(16px) androunded-3xl(24px) for the larger surfaces,rounded-fullfor pills, avatars and lamps. - Sizes: Tailwind's 4px spacing scale. Icons are Lucide,
size-icon-md(16px), orsize-icon-sm(14px) in compact controls. Controls areh-control-sm,h-control-mdorh-control-lg(32, 36, 40px). - Surfaces: one surface utility per element (
card-surface,matte-float,matte-key,matte-well), never rawshadow-*,border-*andbackdrop-blur-*combinations. - Tokens: read a semantic token directly. A component token is only for a value the component decides itself. A value that follows from another token is mixed from it, not picked again.
- No arbitrary values (
text-[11px],rounded-[1.125rem],p-[7px],z-[20]): pick the nearest from the scale (z-20,max-w-4/5,min-w-32). Brackets are for what no scale can hold: acalc(), a grid template, a list of transition properties. - Motion:
ease-out-quintfor what enters or answers a press,ease-sheetfor a sheet or a page that slides. - Phones: every dialog is full screen; only alerts stay small and centred.
- The conversation is the loudest thing on screen; everything around it stays quiet.
Overview
Llama UI is the design system and component library this file describes. Llama Chat, built from it, is the local, private chat client that llama-server bundles, and the native Llama apps (Llama for Mac, Llama for Windows) open it in their own window. It should feel calm and tool-like: neutral greys, plenty of space, and surfaces that read as frosted matte glass rather than shiny plastic. Depth comes from blur, a faint hairline edge and soft shadows, never from strong gradients. Colour is reserved for meaning: the status colours (success, warning, error and info), destructive actions, and the content itself.
There are three layers between a person and a model: the model (the weights, a GGUF
file), the service (the server that runs inference: llama.cpp and llama-server) and the
application (the harness, the UI and UX). The harness, the agentic loop, is its own headless
package, Llama Harness (@ggml-org/llama-harness), next to Llama UI (@ggml-org/llama-ui). Llama UI and Llama Chat are the application
layer. Running all three locally is the point: people choose their model and keep their
conversations and traces, so the UI never assumes a cloud account or hides which model runs.
The service need not be on the same machine: a server can be run by one person and shared by
many (a family, a team, a community), so the UI should make clear which server a model runs on.
This file is the how, written for agents first. Its principles, as rules to apply here:
- Function first. Every element has a job and takes its measurements from a shared role (control heights, corners, icon sizes, type roles); never give one component its own.
- Simple by default, deep on request. The first screen offers a suggested model and sensible defaults; advanced settings (sampling, penalties, system prompts, recipes) sit one step away behind a disclosure, never on the default path. Prefer plain words to terms like GGUF, quantization or speculative decoding, and explain a term where it must appear.
- The content is the loudest thing. The conversation carries the weight and the colour; the interface around it stays neutral and quiet.
- Colour has one job. Greys build everything; status colours show status and nothing else; art colours and cinder are for artwork and the odd brand detail, never for the UI's controls.
- Recognisable, easy to change. Keep the character in the icon, the type and one accent, so an app built on Llama UI can restyle it through semantic and component tokens alone.
- Design lives in code. The CSS is the source of truth; prefer what is generated or
derived over hand-tuned values, and record the reason next to the value. Content pages are
Markdown with Svelte components in them, as
.svx(seeAGENTS.md). - Reuse before adding. An existing role before a new token, an existing component before a new one.
Design tokens come in three tiers, each in its own file under ui/lib/ui/src/lib/styles/tokens/:
-
Primitives (
primitives.css): raw colours named by what they are, such asneutral-900,cool-gray-200,red-550orwhite-a30. Only semantic tokens point at them. Sizes are not primitives: a role holds its own value, and spacing is Tailwind's 4px grid. -
Semantic tokens (
semantic.css): named by purpose, such assurface-glass,fg-muted,line-field,edge-focus,action-primaryorfill-warning, and the only tier that changes between light and dark. Each is either a pick, which names a primitive per theme, or a derivation, which is mixed from another semantic token by a fixed rule and so follows it through every theme with a single declaration:- soft text is the text colour at 70% (
fg-soft); - a hover fill is the fill at 90% (
action-primary-hover); - a status surface is its fill mixed into the page, and a status line its fill mixed into the default line, by a share the theme sets (11% and 34% in light, 16% and 38% in dark);
- a dark shadow is the light one,
elevation-strength(1 in light, 5 in dark) times as dense; - the middle of a matte edge is its side at two thirds, and a lamp's rim and glyph are
its fill mixed 30% towards white and towards black (both in
surfaces.css).
The type, icon size and control height roles live here too and are Tailwind utilities:
text-label,size-icon-md,h-control-md. Corners are Tailwind's ownrounded-smtorounded-4xl, computed from--radiusintheme.css. shadcn's colour names (background,muted-foreground,border...) exist as Tailwind utilities only, each mapped to a semantic token intheme.css. - soft text is the text colour at 70% (
-
Component tokens (
components.css): only what a component decides itself, named[component]-[variant]-[property]-[state], such asbutton-secondary-background-hoverorfloat-panel-background-modal: a value no semantic token has, or one that only this component changes between themes. A component that takes a semantic token as it is reads the semantic token, asbg-(--action-primary)and similar, and gets no alias here.
The front matter's colors, typography and rounded are generated from the primitives, the
semantic tokens and the font stacks by npm run design:tokens in ui/apps/chat, and CI fails
when they are out of date (npm run check:design): change the CSS and run it, never edit them
here. The spacing and components sections and the prose are written by hand; when they
disagree with the CSS, the CSS wins and this file is updated to match. Storybook shows every
tier under Llama UI › Foundations › Colors.
Paper light mode is the base. A semantic token that resolves differently in dark mode has a
-dark twin here, and one that resolves differently in Matte a -matte twin; components list
a -dark entry wherever their colours change. Translucent component
fills are given as they render over the page, so contrast checks see the real colour.
Colors
The palette is achromatic. Hue appears only in the hairline edges (a cool grey at hue 250,
kept almost neutral) and in the status colours, which serve function only: green for success,
amber for warning, red for errors and blue for information, saturated enough to read at a
glance and tuned to text contrast in both modes. Each status has one root colour, its fill (fill-*),
and everything in that status derives from it: the soft surface and line of alerts and badges,
the border of an invalid field, the lamp. Red keeps to the same rule, so there is one red: the
danger button and the solid danger badge are the fill itself (action-danger), solid in light
mode and frosted, translucent matte glass in dark (58% of the fill over the page). Never style a
danger control with its own red (bg-destructive, bg-red-500 …): use the destructive
variant, which AlertDialog.Action takes too, or the status tokens. Cinder (art-cinder, oklch(0.68 0.18 46), #ee6c1e) is the brand's one accent orange: the .cpp of the llama.cpp
logotype, the mark on a detail in the brand pages and the warm pop in artwork. It is not part of the UI palette. Sand (sand-50 to sand-900) is the warm neutral for communications
(posters, boards, the site); the UI itself stays on the greys.
Backgrounds stay neutral as a rule, in the UI and in communications alike: greys and sand
behind, colour only where it serves a purpose, functional (status, a selection, focus) or
creative (a shape, a numeral, an illustration). Artwork takes its colours from the greys, sand and
the art-* primitives: the day sky and the night sky (llama.app's light and dark hero grounds) and
cinder, the accent orange; never from the status colours, and never red.
Light mode has two surface themes, which share everything else (the palette, buttons, fields,
badges, shadows and dark mode): Paper, the default, sets white panels, fields and keys on a
faint grey page (neutral-25); Matte sets grey glass on a white page. Matte is the
theme-matte class on <html>, chosen with the Light surfaces setting (Settings → General) or
the Surface switch in Storybook's toolbar, and only its surface tokens differ: surface-page,
surface-glass, surface-code, field-background and float-panel-background-modal.
- Surfaces:
surface-pagefor the page,surface-mutedfor quiet panels,surface-emphasisandsurface-selectedfor stronger fills and hover or selected rows,surface-raisedfor menus and dialogs,surface-liftedfor a selection raised on a recessed track,surface-pressedfor a key pressed in,surface-glassfor the translucent matte fill,surface-codefor code andsurface-overlaybehind dialogs. - Foreground:
fg-defaultandfg-mutedfor text and icons,fg-soft(the text colour at 70%) for the label of an option that is not the picked one,fg-on-actionandfg-on-dangeron filled buttons,fg-codefor code,fg-dangerfor errors,fg-successfor confirmations such as a copied check,fg-warningfor cautions andfg-infofor neutral notices. - Lines:
line-defaultfor dividers,line-controlfor outlined controls (soft in dark mode),line-fieldfor inputs,line-focusfor focus andline-dangerfor invalid fields. - Edges: the hairlines of matte surfaces, darker at the sides than in the middle (the
token is the side; the surface fades it to two thirds in the middle).
edge-restat rest,edge-focuswhen focus is inside,edge-ambientfor large quiet containers whose edge never changes and, kept even all round, for the user bubble, andedge-raisedfor floating panels. - Actions:
action-primary(near black in light mode, near white in dark mode) and its hover,action-dangerand its hover, andaction-ghost-hover. - Status fills:
fill-success,fill-warning,fill-dangerandfill-infoare each status colour at its brightest, for dots, lamps and other small solid marks. They are brighter than the matchingfg-*text colour, which keeps to a readable lightness (4.5:1 on the soft surfaces), and are a little lighter again in dark mode. - Status surfaces:
surface-success,surface-warning,surface-dangerandsurface-infoare soft tints of each status fill for alerts and soft badges, withline-*-softhairlines to match. They are mixed in oklab so the hue holds over the page. Soft status badges are pills with a small dot in the status fill before the label. - Status:
fill-danger,fill-warning,fill-successandfill-infoare the lamp fills for error or stopped, loading or warning, ready or connected, and informational. A lamp adds a lighter rim and a darker glyph, both mixed from its fill. Known gap: the glyph on the red lamp reaches only about 2:1 and on the green lamp about 3.2:1, below WCAG AA, so the lamp colour must never be the only signal; pair it with a label or tooltip.
Typography
The UI uses Llama Sans for all text, headings included (with the system sans-serif stack as its
fallback), and Llama Mono for code (with the system monospace stack as its fallback). Llama Sans
is one variable font with an optical size axis: the browser sets it from the font-size, so text at
16px gets the UI spacing and kerning, 36px and up the tighter Display setting and text under 16px a
looser Small one, with a blend in between. Nothing needs choosing per element; pin a point only to
show it, with font-variation-settings: 'opsz' 16 or 36. Its line box is 1.25 at every size, centred on the
cap height; large headings set line-height: 1.15 themselves. Llama Mono keeps every character,
its coding ligatures included, in one cell, so code and numbers line up in columns. Type is set through roles that fix size and line height together, and for labels and headings the weight too:
body(16/24): chat messages and the chat input.body-sm(14/20): most text, fields and lists.label(14/20, 500): buttons and field labels.label-sm(12/16, 500): badges and compact labels.caption(12/16): tooltips, hints and metadata.caption-sm(10/14): the smallest text, for tags, compact badges and counts.title(18/28, 600) andheading(24/32, 600): rare headings.display(30/36, 600): the biggest headline, such as the greeting on a new chat.
Body and caption text takes the weight around it, regular unless a parent says otherwise; add
a weight class for anything else (text-body-sm font-semibold). The roles are the whole
scale: no text-sm or text-xs, and no size in brackets. On a phone a row that needs larger
text takes the next role up (text-body-sm max-md:text-body).
When a component lets callers override the size, keep an explicit weight class next to the
role (text-label font-medium) so a size override does not drop the weight. Placeholders
in the chat form are engraved: muted text with a 1px highlight below it in light mode and a
1px shadow above it in dark mode (the engraved utility).
Layout
Spacing follows a 4px grid. Gaps of 4 to 12px (xs to md) dominate inside controls and
lists; 16 to 32px (lg to 2xl) separate sections. Controls are 32, 36 or 40px tall
(h-control-sm, h-control-md, h-control-lg) and icons are 12, 14, 16 or 20px
(size-icon-xs to size-icon-lg). Square boxes use size-N, never a matching h-N w-N
pair, and icons use the size-icon-* roles. Inside buttons, toggles and menu items an icon is
16px (size-icon-md) unless a smaller control calls for less. The chat column is centred with
a generous bottom offset so the floating chat form never covers the last message. The desktop
sidebar is 18rem wide by default, can be resized, and collapses to a 3rem strip.
Safe areas
The page runs under the status bar and over the home indicator (viewport-fit=cover), which
an installed iPhone app always shows and a browser tab mostly does not. The insets are
--safe-area-top and --safe-area-bottom (0 where there is none). Anything pinned to a screen
edge keeps clear of them with the safe-area utilities from src/lib/styles/safe-area.css,
never a hand-written calc():
pt-safe,pb-safe: padding of just the inset, e.g. a sheet that fills the screen.pt-safe-N,pb-safe-N: a spacing step plus the inset, e.g.pt-safe-12for the chat, whose messages start 3rem below the top on a phone.pt-(--page-top): where a page with a title, like Models, Projects or Library, starts on a phone: a row below the corner controls, so the title clears both the menu key and the fade under the status bar.top-safe-N,bottom-safe-N: the same for absolutely or fixedly placed controls, e.g.top-safe-2.5for the expanded chat form's collapse button.top-(--corner-control-top),bottom-(--corner-control-bottom),right-(--corner-control-side): the phone's corner controls, like the menu key and a page's floating key. They follow the insets differently in the installed app (see Corners below).
Corners in the installed app. The status bar and the home indicator set the screen's top
and bottom apart on their own, so a corner control does not add its usual gap on top of their
insets there. At the top it sits just below the status bar (--corner-control-top, the inset
plus 0.25rem). At the bottom the home indicator is a short bar in the middle and the corners
beside it are free, so a bottom corner action ignores its inset instead of stacking above the
indicator. Its gaps from the bottom and the side are always equal, 1.75rem
(--corner-control-bottom is defined from --corner-control-side), which keeps a 56px key
about concentric with an iPhone's rounded screen corner. Only what spans the bottom edge's
middle, like a sheet's last row, clears the inset with pb-safe or bottom-safe-N. In a
browser tab the top keeps 1rem and the bottom its 1.75rem beyond the insets.
A control placed inside a padded box still needs its own top-safe-N: absolute positions
ignore the padding. Dialogs keep clear of the insets on their own (see Dialogs on phones under
Components). At the top, every route on a phone has the status-bar-fade from the root
layout: solid behind the status bar and the row of corner controls under it, fading out down to
--page-top, like the fade under the desktop tab bar, so content scrolls away under them
instead of colliding with the clock or the chat's header. A route needs no fade of its own;
what covers the whole screen, like the expanded chat form, hides it, and it fades out as the
page slides aside for the sidebar. Check
pinned content with the insets simulated (the Safe areas page in
Storybook does this), since they are 0 in a desktop browser.
Elevation & Depth
Depth is built from the matte utilities in ui/lib/ui/src/lib/styles/surfaces.css. Pick the one that
matches the role of the surface instead of combining raw shadows.
matte-glass: large resting surfaces. A translucent fill with a backdrop blur, an inner highlight, a soft drop shadow and a 1px hairline whose sides are slightly darker than its middle. The hairline is drawn on a masked::beforeso it never tints the fill. Focus inside swaps the hairline to the focus edge; the shadow stays the same. Its colours come from slot variables, andchat-form-surface,card-surface,sidebar-surfaceandinset-group-surfacepoint those slots at one component's tokens, so each component needs a single class.matte-floatandmatte-float-modal: menus, popovers, select lists, hover cards and drawers, and dialogs over the page colour. An 86% opaque panel with blur, a faint edge, a rim and a deeper shadow. The edge is a hairline ring baked into the shadow as its last layer (edge-raised). The panel's 1px border is its rim, the inner highlight (highlight-inner) right inside the ring; an inset shadow for the rim would sit inside the border instead, and the 1px between the two would draw a double edge. In light mode dialogs are solid white instead: over the dimmed page a translucent white turns grey.matte-key: buttons, selects, toggles and tabs. The matte finish: on top of the button's own fill it adds a light blur, a fill barely lighter at the top (--key-sheen, 12% white in light mode, 2% in dark) and a small drop shadow. It never changes the button colour and draws no rim of its own: everymatte-keycontrol has a border, and a rim inside it would double the edge. Nothing marks the top edge alone, so the edge looks the same all round, with no gloss.matte-key-solid: primary and danger buttons. The same sheen, fainter (--key-sheen-solid), the same blur, and an even 1px hairline in a shade of the fill drawn inside the button (--key-edge), so the width does not change. The danger key stays matte glass rather than metal: its sheen is almost none (--button-danger-sheen), and in dark mode it is a frosted, translucent red (fill-dangerat 58%) whose edge is a soft tint of the red, never a lighter rim.elevated-control,elevated-surfaceandelevated-raised: cards, floating bars and thumbnails. An elevation token with a hairline ring baked in as its last layer, so the edge melts into the shadow and the shadow deepens in dark mode. The ring is--elevation-ring(black 5%, white 18% in dark), 1px wide. They replace a border plus a shadow; never add a border to them.matte-well: inputs, textareas and checkboxes. A shallow inset shade so fields read as recessed.matte-sheenandmatte-sheen-solid: the user message bubble and badges. An almost invisible top-to-bottom sheen, fainter on solid fills.status-lampandmatte-orbwithorb-danger,orb-warningororb-success: small status dots and the recording button. A near-flat fill barely lighter at the top and an even hairline, with no glow.
The elevation tokens are elevation-surface, elevation-control, elevation-raised,
elevation-inset and elevation-pressed; a tooltip takes elevation-surface. The drop shadows are smooth stacks: a few soft
layers instead of one hard one, the closest the darkest and smallest, each further layer about
twice the offset and blur and fainter, the far ones pulled in with a negative spread, so the
shadow fades out instead of ending in a visible edge. They are the only shadows: Tailwind's
own shadow-sm to shadow-2xl do not follow the theme and are not used. elevation-surface
reaches about 20px, elevation-raised about 38px, and elevation-control is three tight
layers reaching no further than 4px, so buttons and other controls sit close to the page. Two softer steps sit below them: elevation-soft for the chat form,
which floats only a little, and elevation-faint for the user message bubble, which barely
lifts. The keys inside the chat form, the add key and the model selector, take elevation-key:
a 0.5px hairline all round and a 1px drop, since the control elevation's soft
spread blends into the white form; in dark it is the control elevation. Dark mode deepens the same stacks. Tailwind's own shadow-xs to shadow-2xl are
built the same way (tokens/semantic.css), each reaching about as far as Tailwind's default for
it. Every
button carries elevation-control (--button-shadow); a ghost button only while hovered. Overlays behind dialogs are surface-overlay
with a 2px blur.
Shapes
Corners are soft and consistent, on one scale that is computed from a single base radius
(--radius, 10px, in theme.css), so changing the base moves every corner with it: sm
(6px) for small marks and chips, md (8px) for buttons, fields and badges, lg (10px) for
menus and dialogs, xl (14px) for cards, 2xl (16px) for the sidebar and message bubbles,
and 3xl (24px) for the chat form, rising to 4xl (32px) for the chat form on small
screens. Status lamps, avatars and the round icon buttons in the chat form are fully round.
Brand
- Logo: the llama.cpp mark, one inline SVG (
Logo) filled with the current text colour, so it follows the theme with no light and dark copies. It is sized by--size, 1rem by default: 14px in model rows, 16px in the sidebar. Colour it with a text token such asfg-defaultorfg-muted, never a raw colour. - App icons: the mark in
#111111on white, or#fafafaon#111111in dark mode (FAVICON_COLORS,THEME_COLORS). The favicon keeps a 4% margin (FAVICON_PADDING), so the logo spans 96% of it. The maskable and Apple touch icons are white squares with the logo at 67% (the generator's 30% padding); its farthest point reaches 73%, inside the 80% maskable safe zone. The PWA icons are white circles with the logo at 57% (make-icons-circular.js, 85% of the square icon). - Splash screens: white, or
#111111in dark mode, with the logo centred at 24% of the screen's shorter side (SPLASH_PADDING). - The icons and splash screens in
static/are build output ofnpm run build-pwa-assets, drawn fromlogo.svg; change the source or the constants, never the generated files.
Components
-
Buttons: six variants.
default: the solid primary button, including the chat submit button.secondary: the matte glass button, the same fill as the chat form with a faint border, also used for the chat form's Stop and Skip reasoning buttons.tertiary: a faint grey tint with a soft muted border, quieter than secondary and still a visible button; kept for low-emphasis actions that need more presence than ghost.outline: the input fill with aline-controlborder, for secondary actions in dialogs and forms.ghost: no fill until hovered, for icon buttons and quiet actions.destructive: red, for deleting and other irreversible actions.
All except ghost carry
matte-key. Sizes are sm (32px), default (36px) and lg (40px), plus square icon sizes. -
Badges: four variants with the same names:
default,secondary(a light foreground tint),outlineanddestructive, plus the soft status badgessuccess,warning,dangerandinfo. Badges are a fill or a border with the faint matte sheen and no inner rim. -
Toggles: the same corners, type, heights, padding and icon size as buttons, with square
icon-sm,iconandicon-lgsizes for icon-only toggles.defaulthas no fill until hovered;outlineborrows the outline button. Pressed toggles takesurface-selected. -
Button groups: joined buttons share one shadow, drawn by the group, and one border at each seam. A group separator replaces the seam border. Filled buttons draw a seam in their own text colour at 25% instead.
-
Selection on a track: a choice that moves along a track is a recessed track with the selection raised on it, never a row of outlined buttons. The track is
surface-mutedwith the inset shade (elevation-inset); the selection is asurface-liftedchip withelevation-control, and unselected options are softened text (fg-soft) with no fill. Tabs use the same track for their list, with the active tab raised; alinevariant underlines the active tab instead. The chip or underline slides between tabs in 300ms on a plain ease-out, with a stretch of at most 5% along the move that is gone by the time it lands; with reduced motion it jumps instead. The switch is the same track with a raised thumb; switched on, the track takes the primary colour. The message statistics view switcher uses the tabs tokens. -
Toggle groups: a row of keys, like an old stereo's, not a track. At rest the group is a secondary button, the same glass fill, border and shadow as a select trigger, with a
line-controlseam between keys. A pressed key turns white (surface-pressed, the lightest fill in dark mode) and sinks in under a subtle inner shadow (elevation-pressed), its label a pixel lower. While any key is pressed, the others turn grey in light mode (surface-muted) and mute their labels, so the pressed one stands out. A group is as tall as a control of its size. -
Chat form: one
chat-form-surfaceholding the input, attachments and actions. Its fill is the muted grey at 60% in light mode and 75% in dark mode. -
User message bubble: the primary colour at 5% with
matte-sheenand the very faintedge-ambienthairline, so it separates from the page without a visible border. -
Sidebar:
sidebar-surface, the muted fill at 60% with no inner highlight and the ambient edge. On small screens it keeps a plainshadow-mdinstead. -
Inset groups:
inset-group-surfacegathers rows inside a glass panel, like the sidebar footer's Models, MCP and Settings, instead of a rule between them and the rows above. In light mode it is a soft white (surface-raisedat 85%) on the grey glass; in dark mode a raised fill would be darker than the glass, so it is a 3% tint of the text colour. It takes the sidebar's ambient edge, no inner highlight and only a faint, wide shade under it (0 2px 8px -4px, 8% black in light, 40% in dark), with the card radius. It reads as part of the panel: no tight contact shadow and no solid border, which make a button, and no inset shade, which makes a field. Unlike the othermatte-glasssurfaces it does not blur: it already sits on glass, and a second backdrop blur re-paints the panel's fill behind it and turns the group visibly lighter, whatever its own fill. -
Selected rows:
selected-row-surfacemarks the row a sidebar list is on: a conversation, a project or a top action, on desktop and on a phone. It is a 5% tint of the text colour, under hover's 10%, with the inset group's edge and shade, so the selected row reads as picked out rather than hovered, in both modes. It keeps the row's own radius, and hover and press fills still paint over it. -
Fields:
matte-wellwith a 1pxline-fieldoutline and the standard 3px focus ring. -
Floating panels: every menu, popover, select list, hover card and drawer uses
matte-float; dialogs usematte-float-modal. -
Chat header on phones: a chat shows its title and, under it, its working directory centred between the menu key and the context gauge, on the top fade. The directory is the folder's name (the full path is its title) and tapping it opens the directory picker, which on a phone is a drawer. A new chat has no header yet: it offers "Set working directory" in the + menu instead, and no chip sits under the form on a phone. The header steps aside with the corner keys while the expanded chat form covers the screen.
-
Dialogs on phones: below
mdevery dialog is full screen, whatever size it takes on a desktop. The sharedDialog.Contentdoes this, so a dialog sets only its desktop size and never its own phone size. Its top and bottom borders take the height of the status bar and the home indicator, transparent, so the content keeps clear of both with no change to its padding (see Safe areas). A dialog that draws its own full-screen content instead, like the attachment preview, keeps clear withpt-safe,pb-safeandtop-safe-Nitself. Alerts (AlertDialog: confirmations, renames, errors) are the exception: a short question with a button or two, they stay small and centred on a phone too. A place the sidebar opens, like Settings, MCP Servers or Servers, is no dialog on a phone at all but a page at its own route (#/settings,#/mcp-servers,#/servers), with the menu key and a title like any phone page; a desktop or tablet keeps it a dialog over the page it was opened from. -
Glass keys: a round key that floats over the page and lets it show through, like the menu and context gauge keys at a phone's top corners and the key back to a chat's bottom, is a
GlassKey(GLASS_KEY_CLASS): white glass (surface-glass), a faintline-defaulthairline, andmatte-key's blur and saturation with its small shadow, so what scrolls under it softens instead of showing sharp behind the icon. 40px at the corners, 36px over the page; its icon is a 2px-stroke glyph infg-defaulton the 24px icon grid, the context gauge's dial drawn the same way. Don't rebuild one from a ghostButtonplus classes: the button's own blur overridesmatte-key's. -
A page's main action on phones: below
mdthe action that adds to a page, like New project or New prompt, leaves the page's header for aFloatingActionButton: a round 56pxdefault(primary) key with a large shadow and only a+icon, pinned in the bottom-right corner 1.75rem from both edges (--corner-control-side,--corner-control-bottom; in the installed app beside the home indicator rather than above it, see Safe areas), where the thumb rests. It is portalled to the body, so its place on screen never depends on the page's box, and slides with the page when the sidebar opens, by the same offset and ease (--sidebar-reveal-x), taking no taps while the page is aside. The page keeps room under its last row (max-md:pb-safe-24) so the key never covers it. One per page, only for its main action; frommdup the action stays a button in the header. -
Voice clips: a recorded clip is a round primary play key, its waveform and the time in
caption: bars in the muted foreground at 40%, the played part in the primary colour, so the clip stays neutral in both modes. The live waveform while recording is as tall as the input's line (24px), the newest level on the right, and holds still with reduced motion. A voice message's transcript opens under the clip on hover, keyboard focus or a tap, and stays in the DOM for screen readers while closed. -
Copy feedback: a copy action turns its icon into a
fg-successcheck for 1.5s (COPY_FEEDBACK_RESET_MS), only when the copy succeeded, then back. The icon pops in with a 150ms fade and scale, skipped on first render and with reduced motion. -
Tooltips: the inverse of the page, the primary colour with a faint inner rim.
-
Status: server, model and MCP connection states use the status lamps, never coloured text alone.
Do's and Don'ts
- Do keep fills flat. Gradients, if any, must be barely perceptible; a visible gradient makes the UI look like a 2000s interface.
- Do keep edges faint. At rest a hairline should blend in; on focus it should be noticeable but never strong.
- Do keep the shadow the same on focus. Focus is shown by the edge and the focus ring, not by lifting the surface.
- Do use the status colours only for status. Buttons keep their neutral colours, and the submit button stays black in light mode and white in dark mode.
- Do read a semantic token directly wherever a component takes it as it is. Add a component token only for a value the component decides itself, never as an alias of a semantic token.
- Do add a new colour as a primitive first, then give it a semantic name before any component uses it.
- Do give motion a job: show where something went or what changed, like a selection sliding to its new place. A barely visible stretch along a move, or a settle past the target by a pixel or two, can make that feel physical; keep it under about 5% and 300ms, and drop it entirely with reduced motion. Never let a bounce carry something past the edge of its container, such as a tab indicator past its track; there, slide without one. If you notice the effect itself rather than the change it shows, it is too much.
- Do reach for the surface utility that matches the component instead of adding raw
shadow-*,border-*andbackdrop-blur-*combinations. - Don't use Tailwind's palette classes (
text-amber-500,bg-green-500/10) for meaning in new code; use the semantic tokens. Some older components still do and are being moved over. - Don't introduce new accent hues (such as a blue submit button) without a design decision.
- Don't put white rims or highlights on large ambient containers such as the sidebar.
- Don't animate for decoration: no motion on things that did not change, no bounces on entrances, and nothing that delays the user.
- Don't stack two glass surfaces with blur directly on top of each other.
- Don't give a dialog its own size on a phone: every dialog is full screen there. Only alerts stay small and centred.
- Don't put more than one floating key on a phone page, or use one for anything but the page's main action.
- Don't lift a bottom corner action above the home indicator's inset in the installed app,
or give it a different gap from the bottom than from the side: the corners beside the
indicator are free, and
--corner-control-bottomand--corner-control-sidekeep the two gaps equal.