\n \n \n \n\n",[],[249,251],"ebe3384f-c433-45f5-bdcd-ac477f85e588","favicon.svg","01a068eb-ca37-739b-bb70-fe32b88eb1de",["D",1788466678331],"\n",[],[249,259],"c1ed8fb9-35ee-4a84-9896-801031712dd5","VoiceButton.tsx","a8ef1b14-a7d4-11f1-9f9d-1607ee4eb77e","01a068eb-ca37-739b-bb70-fbb512b8ccd8",104,["D",1788473397276],17,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\nimport { useEffect, useRef, useState } from \"https://esm.sh/react@18.2.0\";\n\n// Hands-free voice input through the browser's speech recognizer (Chrome and Safari).\n// A sentence is sent when the recognizer marks it final, or after `silenceMs` without new words.\n// If more words arrive within `continueMs` of a send, the sentence was cut short: `onContinue` gets the\n// whole sentence so the caller can undo the short one and resend. The recognizer keeps running while\n// the agent works; a sentence finished during a run is sent when the run ends.\n// Push-to-talk is the other mode: `hold()` listens until `release()`, then sends everything heard, with no\n// silence cutoff and no continuation. Holding turns hands-free off.\nconst norm = (s: string) => s.toLowerCase().replace(/[^a-z0-9 ]/g, \"\").replace(/\\s+/g, \" \").trim();\nconst wordsOf = (s: string) => s.split(/\\s+/).filter(Boolean);\n\nexport function useVoice({ silenceMs, continueMs, busy, onFinal, onContinue }: {\n silenceMs: number;\n continueMs: number;\n busy: boolean;\n onFinal: (text: string) => void;\n onContinue: (text: string) => void;\n}) {\n const [on, setOn] = useState(false);\n const [holding, setHolding] = useState(false);\n const ptt = useRef(false); // push-to-talk in progress: no silence cutoff, send on release\n const [listening, setListening] = useState(false);\n const [interim, setInterim] = useState(\"\");\n const [status, setStatus] = useState(\"\");\n const onRef = useRef(false);\n const busyRef = useRef(busy);\n busyRef.current = busy;\n const rec = useRef(null);\n const timer = useRef(undefined);\n const consumed = useRef(0); // recognizer results that belong to sentences already dealt with\n const sent = useRef<{ said: string; at: number } | null>(null); // the last sentence sent from this session\n const draft = useRef<{ said: string; cont: boolean } | null>(null);\n const pending = useRef(null); // a sentence that finished while the agent was busy\n const lastWordAt = useRef(0);\n const SR = typeof window === \"undefined\" ? null : (window as any).SpeechRecognition || (window as any).webkitSpeechRecognition;\n const supported = !!SR;\n\n function commit(why: string) {\n const d = draft.current;\n draft.current = null;\n clearTimeout(timer.current);\n setInterim(\"\");\n if (!d || !d.said) return;\n const since = Math.round(performance.now() - lastWordAt.current);\n sent.current = { said: d.said, at: performance.now() };\n setStatus(`${why}, ${since} ms after the last word`);\n console.log(`[voice] ${d.cont ? \"continuation\" : busyRef.current ? \"queued (agent busy)\" : \"send\"}: \"${d.said}\" (${why}, ${since} ms after the last word)`);\n if (d.cont) onContinue(d.said);\n else if (busyRef.current) pending.current = d.said;\n else onFinal(d.said);\n }\n\n function start() {\n if (!SR || rec.current) return;\n const r = new SR();\n r.lang = \"en-US\";\n r.interimResults = true;\n r.continuous = true;\n consumed.current = 0; sent.current = null; draft.current = null;\n r.onstart = () => { setListening(true); setStatus(\"listening\"); console.log(\"[voice] recognizer started\"); };\n r.onresult = (e: any) => {\n const results: any[] = Array.from(e.results);\n const now = performance.now();\n if (ptt.current) {\n // Holding the key: just keep collecting words. Release sends them.\n const said = results.map((x) => x[0].transcript).join(\"\").trim();\n lastWordAt.current = now;\n draft.current = { said, cont: false };\n setInterim(said);\n return;\n }\n const s = sent.current;\n if (s && now - s.at >= continueMs) {\n // The last sentence is settled. Everything final up to now belongs to it.\n let n = 0; while (n < results.length && results[n].isFinal) n++;\n consumed.current = Math.max(consumed.current, n);\n sent.current = null;\n }\n let said = results.slice(consumed.current).map((x) => x[0].transcript).join(\"\").trim();\n // The settled sentence's last result can still be interim; drop its words.\n if (!sent.current && s && norm(said).startsWith(norm(s.said))) said = wordsOf(said).slice(wordsOf(s.said).length).join(\" \");\n // The recognizer re-emits a sent sentence with final casing and punctuation; that is not new words.\n if (!said || (sent.current && norm(said) === norm(sent.current.said))) return;\n const cont = !!sent.current; // words arrived soon after a send: this is the same sentence, longer\n const final = results.slice(consumed.current).every((x) => x.isFinal);\n lastWordAt.current = now;\n draft.current = { said, cont };\n setInterim(said);\n clearTimeout(timer.current);\n if (final) commit(cont ? \"sentence continued\" : \"sentence ended\");\n else timer.current = setTimeout(() => commit(cont ? \"sentence continued\" : `${silenceMs} ms of silence`), silenceMs);\n };\n r.onerror = (e: any) => { if (e.error !== \"aborted\" && e.error !== \"no-speech\") setStatus(`microphone: ${e.error}`); };\n r.onend = () => {\n rec.current = null;\n setListening(false);\n // After a push-to-talk release the recognizer delivers its last words, then ends: send them now.\n if (draft.current) commit(ptt.current ? \"key released\" : \"recognizer stopped\");\n if (ptt.current) { ptt.current = false; setHolding(false); setStatus(\"\"); }\n else if (onRef.current) start();\n };\n rec.current = r;\n r.start();\n }\n\n function stop() {\n onRef.current = false; setOn(false);\n clearTimeout(timer.current); draft.current = null; pending.current = null; setInterim(\"\");\n try { rec.current?.abort(); } catch { /* ignore */ }\n rec.current = null; setListening(false); setStatus(\"\");\n }\n\n function toggle() {\n if (on) stop();\n else { onRef.current = true; setOn(true); start(); }\n }\n\n // Push-to-talk: listen from key down to key up, then send what was heard.\n function hold() {\n if (!SR || ptt.current) return;\n if (onRef.current) stop(); // hands-free is off from here on; the mic button turns it back on\n ptt.current = true; setHolding(true);\n setStatus(\"listening while the key is held\");\n start();\n }\n function release() {\n if (!ptt.current) return;\n const r = rec.current;\n if (!r) { ptt.current = false; setHolding(false); return; }\n setStatus(\"sending\");\n try { r.stop(); } catch { r.onend?.(); } // stop() delivers the final words, then fires onend, which sends\n }\n\n // The agent finished: send what was said meanwhile.\n useEffect(() => {\n if (!busy && pending.current) { const text = pending.current; pending.current = null; onFinal(text); }\n }, [busy]);\n\n return { supported, on, holding, listening, interim, status, toggle, hold, release, start: () => { onRef.current = true; setOn(true); start(); } };\n}\n",[],[249,275,266],"components","ff8bd5aa-c0fe-4384-a0ed-47ccfc86d1e4","Overlay.tsx","fa4b1250-a7d5-11f1-9310-1607ee4eb77e",103,["D",1788473381480],18,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\nimport { useEffect, useState } from \"https://esm.sh/react@18.2.0\";\n\nexport type Toast = {\n id: string;\n at: number;\n kind: \"said\" | \"result\" | \"error\";\n text: string;\n note?: string;\n diffs?: { path: string; added: string[]; removed: number }[];\n stats?: { turns: number; toolCalls: number; tokens: number; tps: number; totalMs: number };\n};\n\n// Lifetimes in ms. Each newer card above a card halves its remaining life, so a burst of activity clears fast.\nconst LIFE: Record = { said: 5000, result: 7000, error: 10000 };\nconst FADE = 1200;\nconst ms = (n: number) => n >= 1000 ? `${(n / 1000).toFixed(2)} s` : `${Math.round(n)} ms`;\n\n// Everything drawn over the page: the mic, and a strip along the bottom, never taller than a tenth of the\n// screen, holding one line each for the words being heard, what the agent is doing, and what was said and\n// what changed. Lines fade out; new ones push old ones off the top.\nexport function Overlay({ toasts, interim, live, mic, typed, onTyped, onSubmit, onExpire }: {\n toasts: Toast[];\n interim: string;\n live: { text: string; tools: string[]; turns: number } | null;\n mic: { supported: boolean; on: boolean; holding: boolean; listening: boolean; status: string; toggle: () => void };\n typed: string | null;\n onTyped: (t: string | null) => void;\n onSubmit: (text: string) => void;\n onExpire: (id: string) => void;\n}) {\n const [now, setNow] = useState(performance.now());\n useEffect(() => {\n const t = setInterval(() => setNow(performance.now()), 100);\n return () => clearInterval(t);\n }, []);\n const life = (i: number) => LIFE[toasts[i].kind] / 2 ** (toasts.length - 1 - i);\n useEffect(() => {\n toasts.forEach((t, i) => { if (now - t.at > life(i)) onExpire(t.id); });\n }, [now]);\n const opacity = (i: number) => Math.max(0, Math.min(1, (life(i) - (now - toasts[i].at)) / FADE));\n\n const line = \"w-full max-w-3xl truncate rounded-lg px-3 py-1 text-[13px] leading-5 backdrop-blur\";\n\n return (\n <>\n \n {mic.listening ? \"●\" : \"🎙\"}\n {mic.holding && Space held}\n \n\n

\n {toasts.map((t, i) => (\n

\n {t.kind === \"said\" && (\n

\n {t.text}\n {t.note && {t.note}}\n

\n )}\n {t.kind === \"result\" && (\n

\n {t.text}\n {t.diffs?.map((d) => {d.path} +{d.added.length} −{d.removed})}\n {t.stats && (\n \n {t.stats.tps > 0 && `${t.stats.tps.toLocaleString()} tok/s · `}{t.stats.tokens.toLocaleString()} tok · {t.stats.turns}t/{t.stats.toolCalls}c · {ms(t.stats.totalMs)}\n \n )}\n

\n )}\n {t.kind === \"error\" &&

{t.text}

}\n

\n ))}\n\n {live && (\n

\n {live.tools.length ? live.tools.at(-1) : \"thinking…\"}\n {live.text && {live.text}}\n

\n )}\n\n {/* The words being heard: one line, pinned to its right end, so the newest words stay visible and the oldest slide off the left. */}\n {interim && (\n

\n {interim}\n

\n )}\n

\n\n {typed !== null && (\n

\n onTyped((e.target as HTMLInputElement).value)}\n onKeyDown={(e) => {\n if (e.key === \"Enter\") onSubmit(typed);\n if (e.key === \"Escape\") onTyped(null);\n }}\n placeholder=\"Say or type what to build or change\"\n className=\"pointer-events-auto w-full max-w-3xl rounded-lg border border-gray-300 bg-white px-3 py-1.5 text-base shadow-lg focus:outline-none focus:ring-2 focus:ring-blue-500\"\n />\n

\n )}\n \n );\n}\n",[],[249,275,277],"b8f2fd75-810c-49fa-9337-062886f5b4c0","App.tsx","01a068eb-ca36-70ad-ae0b-9ba9da2d93cc",["D",1788535911033],36,"/** @jsxImportSource https://esm.sh/react@18.2.0 */\nimport { useEffect, useRef, useState } from \"https://esm.sh/react@18.2.0\";\nimport { useVoice } from \"./VoiceButton.tsx\";\nimport { Overlay } from \"./Overlay.tsx\";\nimport type { Toast } from \"./Overlay.tsx\";\n\n// Voice-first: the page being built fills the screen. What you say appears at the bottom and fades;\n// what changed, and how fast, appears the same way after each run. Typing any key opens a text box.\nconst PARAMS = new URLSearchParams(location.search);\n// Both go to Cerebras directly, so the val only needs CEREBRAS_API_KEY. ?model= overrides with another\n// cerebras:, or any OpenRouter slug (optionally \"@provider\") if OPENROUTER_API_KEY is set.\nconst MODELS = [\n { id: \"cerebras:gpt-oss-120b\", label: \"GPT-OSS 120B\", note: \"OpenAI\" },\n { id: \"cerebras:qwen-3.8-27b\", label: \"Qwen 3.8 27B\", note: \"Alibaba\" },\n];\nconst INITIAL_MODEL = PARAMS.get(\"model\") || MODELS[0].id;\nconst SILENCE_MS = Number(PARAMS.get(\"silence\")) || 1000;\nconst CONTINUE_MS = Number(PARAMS.get(\"continue\")) || 1500; // words this soon after a send extend that sentence\nconst AUTO_VOICE = PARAMS.get(\"voice\") !== \"0\";\n\ntype Snapshot = { files: Record; messages: unknown[] };\ntype Run = { id: number; controller: AbortController; before: Snapshot; saidToast: string; resultToast?: string };\n\nfunction lineDiff(before: string, after: string) {\n const a = before.split(\"\\n\"), b = after.split(\"\\n\");\n const countA = new Map(); for (const l of a) countA.set(l, (countA.get(l) ?? 0) + 1);\n const added: string[] = []; const remaining = new Map(countA);\n for (const l of b) { const n = remaining.get(l) ?? 0; if (n > 0) remaining.set(l, n - 1); else added.push(l); }\n const removed = [...remaining.values()].reduce((s, n) => s + n, 0);\n return { added, removed };\n}\n\nexport function App() {\n const [messages, setMessages] = useState([]);\n const [files, setFiles] = useState>({});\n const [loading, setLoading] = useState(false);\n const [model, setModel] = useState(INITIAL_MODEL);\n const modelRef = useRef(model);\n modelRef.current = model;\n const [toasts, setToasts] = useState([]);\n const [typed, setTyped] = useState(null);\n const [live, setLive] = useState<{ text: string; tools: string[]; turns: number } | null>(null);\n const iframeRef = useRef(null);\n const filesRef = useRef(files);\n filesRef.current = files;\n const messagesRef = useRef(messages);\n messagesRef.current = messages;\n const run = useRef(null); // the latest run, kept so a cut-short sentence can undo it\n const runSeq = useRef(0);\n\n function toast(t: Omit): string {\n const id = crypto.randomUUID();\n setToasts((prev) => [...prev.slice(-3), { ...t, id, at: performance.now() }]);\n return id;\n }\n const dropToast = (id?: string) => { if (id) setToasts((prev) => prev.filter((t) => t.id !== id)); };\n const editToast = (id: string, patch: Partial) => setToasts((prev) => prev.map((t) => t.id === id ? { ...t, ...patch, at: performance.now() } : t));\n\n // A fresh document each time, so the page's own top-level declarations don't collide with the last render.\n function render(html: string) {\n if (iframeRef.current) iframeRef.current.srcdoc = html;\n }\n\n function restore(s: Snapshot) {\n setFiles(s.files); filesRef.current = s.files;\n setMessages(s.messages); messagesRef.current = s.messages;\n render(s.files[\"index.html\"] ?? \"\");\n }\n\n async function send(instruction: string, saidToast: string) {\n const id = ++runSeq.current;\n const controller = new AbortController();\n const before: Snapshot = { files: filesRef.current, messages: messagesRef.current };\n run.current = { id, controller, before, saidToast };\n setLoading(true);\n setLive({ text: \"\", tools: [], turns: 0 });\n const startedAt = performance.now();\n let tokens = 0, decodeMs = 0;\n try {\n const res = await fetch(\"/api/agent\", {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify({ messages: before.messages, files: before.files, instruction, model: modelRef.current }),\n signal: controller.signal,\n });\n if (!res.ok || !res.body) throw new Error((await res.json().catch(() => ({})))?.error ?? `Request failed (${res.status})`);\n const reader = res.body.getReader();\n const decoder = new TextDecoder();\n let buffer = \"\";\n while (true) {\n const { done, value } = await reader.read();\n if (done) break;\n buffer += decoder.decode(value, { stream: true });\n const frames = buffer.split(\"\\n\\n\");\n buffer = frames.pop() ?? \"\";\n for (const frame of frames) {\n const line = frame.split(\"\\n\").find((l) => l.startsWith(\"data:\"));\n if (!line) continue;\n const event = JSON.parse(line.slice(5).trim());\n switch (event.type) {\n case \"text\": setLive((l) => l && { ...l, text: l.text + event.text }); break;\n case \"tool\": setLive((l) => l && { ...l, tools: [...l.tools, `${event.name} ${event.args}`] }); break;\n case \"turn\": tokens += event.stats.completionTokens; if (event.stats.tokensPerSecond > 0) decodeMs += event.stats.completionTokens / event.stats.tokensPerSecond * 1000; setLive((l) => l && { ...l, turns: event.stats.index }); break;\n case \"files\": setFiles(event.files); if (event.files[\"index.html\"] !== undefined) render(event.files[\"index.html\"]); break;\n case \"done\": {\n setMessages(event.messages);\n setFiles(event.files);\n const changed = Object.keys(event.files).filter((p) => event.files[p] !== before.files[p]);\n const diffs = changed.map((p) => ({ path: p, ...lineDiff(before.files[p] ?? \"\", event.files[p]) }));\n const resultToast = toast({ kind: \"result\", text: event.messages.at(-1)?.content ?? \"\", diffs, stats: { turns: event.totals.turns, toolCalls: event.totals.toolCalls, tokens, tps: decodeMs >= 100 ? Math.round(tokens / (decodeMs / 1000)) : 0, totalMs: performance.now() - startedAt } });\n if (run.current?.id === id) run.current.resultToast = resultToast;\n break;\n }\n case \"error\": throw new Error(event.message);\n }\n }\n }\n } catch (err) {\n if ((err as Error).name !== \"AbortError\") toast({ kind: \"error\", text: err instanceof Error ? err.message : String(err) });\n } finally {\n if (run.current?.id === id) { setLoading(false); setLive(null); }\n }\n }\n\n function say(text: string) {\n send(text, toast({ kind: \"said\", text }));\n }\n\n // The sentence kept going after it was sent: undo that run and send the whole sentence.\n function sayAgain(text: string) {\n const r = run.current;\n if (!r) return say(text);\n r.controller.abort();\n restore(r.before);\n dropToast(r.resultToast);\n editToast(r.saidToast, { text, note: \"kept talking, so the first version was undone and resent\" });\n send(text, r.saidToast);\n }\n\n const voice = useVoice({ silenceMs: SILENCE_MS, continueMs: CONTINUE_MS, busy: loading, onFinal: say, onContinue: sayAgain });\n\n useEffect(() => { if (AUTO_VOICE && voice.supported) voice.start(); }, []);\n\n // Hold Space to talk, release to send. Type anything else to get a text box; Enter sends, Escape closes.\n useEffect(() => {\n const onKey = (e: KeyboardEvent) => {\n if (typed !== null || e.metaKey || e.ctrlKey || e.altKey) return;\n if (e.key === \" \") { e.preventDefault(); if (!e.repeat) voice.hold(); return; }\n if (e.key.length === 1) { setTyped(e.key); e.preventDefault(); }\n };\n const onKeyUp = (e: KeyboardEvent) => { if (e.key === \" \" && typed === null) { e.preventDefault(); voice.release(); } };\n window.addEventListener(\"keydown\", onKey);\n window.addEventListener(\"keyup\", onKeyUp);\n return () => { window.removeEventListener(\"keydown\", onKey); window.removeEventListener(\"keyup\", onKeyUp); };\n }, [typed]);\n\n return (\n

History · README.md