Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

chat speed is decode speed with reasoning, context shown is the running model's, Models shows Unload while loaded, VRAM reserve floor, default text size 14, installer branding, icon

Niko12210db5d17102ad
diff --git a/apps/desktop/scripts/make-icons.mjs b/apps/desktop/scripts/make-icons.mjs
new file mode 100644
index 000000000..411e4b97b
--- /dev/null
+++ b/apps/desktop/scripts/make-icons.mjs
@@ -0,0 +1,78 @@
+#!/usr/bin/env node
+/**
+ * The app and installer icon, refined from the existing Strata mark (apps/host/public/brand/strata-mark-*.svg and the current
+ * icons/icon.png): the same three rounded bars on the same dark tile, in the same colours. Nothing is redrawn. What changes is
+ * how it is cut for small sizes: at 16-48 px the bars are snapped to whole pixels and made a little bolder and wider so they
+ * stay crisp instead of blurring; from 64 px up the geometry is the mark's own proportions.
+ *
+ * node apps/desktop/scripts/make-icons.mjs [--out <folder>] writes icon-<size>.png for every size and icon.ico
+ */
+import { existsSync, mkdirSync, writeFileSync } from 'node:fs';
+import { dirname, join, resolve } from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+
+const args = Object.fromEntries(process.argv.slice(2).reduce((acc, a, i, all) => (a.startsWith('--') ? [...acc, [a.slice(2), all[i + 1]]] : acc), []));
+const here = dirname(fileURLToPath(import.meta.url));
+const host = resolve(here, '../../host');
+const out = resolve(args.out ?? join(here, '../src-tauri/icons'));
+mkdirSync(out, { recursive: true });
+const { chromium } = await import(pathToFileURL(join(host, 'node_modules/playwright-core/index.mjs')).href).catch(() => import('playwright-core'));
+const CHROME = process.env.CHROME ?? ['C:/Program Files/Google/Chrome/Application/chrome.exe', '/usr/bin/google-chrome'].find(existsSync);
+
+const TILE = '#161a1d', LIGHT = '#eef1f3', GREEN = '#34d399'; // the colours of the current icon.png
+
+/** Integer geometry for the small sizes: bars on whole pixels, centred. Returns [{x,y,w,h,fill}]. */
+function hinted(size) {
+ const bar = Math.max(2, Math.round(size * 0.14)), gap = Math.max(2, Math.round(size * 0.1)), mid = bar + 1;
+ const total = 2 * bar + mid + 2 * gap, top = Math.floor((size - total) / 2);
+ const even = w => (w % 2 === size % 2 ? w : w + 1);
+ const wide = even(Math.round(size * 0.5)), long = even(Math.round(size * 0.76));
+ return [
+ { x: (size - wide) / 2, y: top, w: wide, h: bar, fill: LIGHT },
+ { x: (size - long) / 2, y: top + bar + gap, w: long, h: mid, fill: GREEN },
+ { x: (size - wide) / 2, y: top + bar + gap + mid + gap, w: wide, h: bar, fill: LIGHT },
+ ];
+}
+/** The mark's own proportions (from icon.png at 512): bars 34% / 50% / 34% wide, 9% high, on a tile with 21.5% corners. */
+function proportional(size) {
+ const k = size / 512;
+ return [
+ { x: 170 * k, y: 156 * k, w: 173 * k, h: 46 * k, fill: LIGHT },
+ { x: 128 * k, y: 232 * k, w: 256 * k, h: 48 * k, fill: GREEN },
+ { x: 170 * k, y: 309 * k, w: 173 * k, h: 46 * k, fill: LIGHT },
+ ];
+}
+function svg(size) {
+ const small = size <= 48;
+ const bars = (small ? hinted(size) : proportional(size));
+ const radius = small ? Math.round(size * 0.22) : size * 0.215;
+ const rects = bars.map(b => `<rect x="${b.x}" y="${b.y}" width="${b.w}" height="${b.h}" rx="${small ? Math.min(b.h / 2, 1.5 + size / 32) : b.h / 2}" fill="${b.fill}"${small ? ' shape-rendering="crispEdges"' : ''}/>`).join('');
+ return `<svg xmlns="http://www.w3.org/2000/svg" width="${size}" height="${size}" viewBox="0 0 ${size} ${size}"><rect width="${size}" height="${size}" rx="${radius}" fill="${TILE}"/>${rects}</svg>`;
+}
+
+const SIZES = [16, 20, 24, 32, 40, 48, 64, 96, 128, 256, 512, 1024];
+const browser = await chromium.launch({ executablePath: CHROME, headless: true });
+const png = {};
+for (const size of SIZES) {
+ const page = await browser.newPage({ viewport: { width: size, height: size }, deviceScaleFactor: 1 });
+ await page.setContent(`<!doctype html><body style="margin:0;background:transparent">${svg(size)}</body>`);
+ png[size] = await page.screenshot({ omitBackground: true, clip: { x: 0, y: 0, width: size, height: size } });
+ await page.close();
+ writeFileSync(join(out, `icon-${size}.png`), png[size]);
+}
+await browser.close();
+
+// The .ico: one PNG image per size (16-256), the sizes Windows asks for at 100-250% scaling.
+const ICO = [16, 20, 24, 32, 40, 48, 64, 96, 128, 256];
+const head = Buffer.alloc(6); head.writeUInt16LE(0, 0); head.writeUInt16LE(1, 2); head.writeUInt16LE(ICO.length, 4);
+let offset = 6 + 16 * ICO.length;
+const entries = [], blobs = [];
+for (const size of ICO) {
+ const data = png[size], e = Buffer.alloc(16);
+ e.writeUInt8(size === 256 ? 0 : size, 0); e.writeUInt8(size === 256 ? 0 : size, 1); e.writeUInt8(0, 2); e.writeUInt8(0, 3);
+ e.writeUInt16LE(1, 4); e.writeUInt16LE(32, 6); e.writeUInt32LE(data.length, 8); e.writeUInt32LE(offset, 12);
+ offset += data.length; entries.push(e); blobs.push(data);
+}
+writeFileSync(join(out, 'icon.ico'), Buffer.concat([head, ...entries, ...blobs]));
+writeFileSync(join(out, 'icon.png'), png[512]);
+console.log('icons in', out, ICO.join(','));
diff --git a/apps/desktop/scripts/make-installer-images.mjs b/apps/desktop/scripts/make-installer-images.mjs
new file mode 100644
index 000000000..f12d6cc15
--- /dev/null
+++ b/apps/desktop/scripts/make-installer-images.mjs
@@ -0,0 +1,66 @@
+#!/usr/bin/env node
+/**
+ * The installer's pictures, drawn from the app's own design: the light surface and ink colours of the app (src/styles tokens), the
+ * brand lockup and mark files, GT Walsheim Bold for the words (baked into the pictures, because NSIS cannot use a web font).
+ *
+ * node apps/desktop/scripts/make-installer-images.mjs [--out <folder>] [--preview <folder>]
+ *
+ * Writes sidebar.bmp (164x314, the welcome and finish pages) and header.bmp (150x57, the other pages) as 24-bit BMPs, and with
+ * --preview the same as PNGs at 3x.
+ */
+import { spawnSync } from 'node:child_process';
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { dirname, join, resolve } from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+import { tmpdir } from 'node:os';
+
+const args = Object.fromEntries(process.argv.slice(2).reduce((acc, a, i, all) => (a.startsWith('--') ? [...acc, [a.slice(2), all[i + 1]]] : acc), []));
+const here = dirname(fileURLToPath(import.meta.url));
+const host = resolve(here, '../../host');
+const out = resolve(args.out ?? join(here, '../src-tauri/nsis'));
+mkdirSync(out, { recursive: true });
+const { chromium } = await import(pathToFileURL(join(host, 'node_modules/playwright-core/index.mjs')).href).catch(() => import('playwright-core'));
+const CHROME = process.env.CHROME ?? ['C:/Program Files/Google/Chrome/Application/chrome.exe', '/usr/bin/google-chrome'].find(existsSync);
+const font = pathToFileURL(join(host, 'public/fonts/GTWalsheimPro-Bold.woff2')).href;
+const fontMedium = pathToFileURL(join(host, 'public/fonts/GTWalsheimPro-Medium.woff2')).href;
+const lockup = readFileSync(join(host, 'public/brand/strata-lockup-light.svg'), 'utf8');
+const mark = readFileSync(join(host, 'public/brand/strata-mark-light.svg'), 'utf8');
+
+// the app's light tokens
+const SURFACE = '#f9fafb', INK = '#171b1e', MUTED = '#5b646b', ACCENT = '#10b981', LINE = '#e3e8ea';
+const css = `@font-face{font-family:GT;font-weight:700;src:url(${font})}@font-face{font-family:GT;font-weight:500;src:url(${fontMedium})}
+*{box-sizing:border-box}body{margin:0;font-family:GT,'Segoe UI',sans-serif;color:${INK};background:${SURFACE}}`;
+
+const pages = {
+ sidebar: { w: 164, h: 314, html: `<div style="width:164px;height:314px;padding:22px 18px;display:flex;flex-direction:column;background:linear-gradient(180deg,#f9fafb 0%,#eef6f2 100%);border-right:1px solid ${LINE}">
+ <div style="height:26px">${lockup.replace('<svg', '<svg style="height:26px;width:auto;display:block"')}</div>
+ <div style="margin-top:auto">
+ <div style="font-weight:700;font-size:22px;line-height:1.12;letter-spacing:-.01em">State of<br>the art.<br>At home.</div>
+ <div style="margin-top:10px;font-weight:500;font-size:11px;line-height:1.5;color:${MUTED}">AI on your own computer.<br>Private. Fast. Easy.</div>
+ <div style="margin-top:16px;height:3px;width:34px;border-radius:2px;background:${ACCENT}"></div>
+ </div></div>` },
+ header: { w: 150, h: 57, html: `<div style="width:150px;height:57px;display:flex;align-items:center;justify-content:flex-end;padding-right:12px;background:${SURFACE}">${lockup.replace('<svg', '<svg style="height:22px;width:auto;display:block"')}</div>` },
+};
+
+const browser = await chromium.launch({ executablePath: CHROME, headless: true });
+const work = join(tmpdir(), 'strata-installer-images'); mkdirSync(work, { recursive: true });
+for (const [name, page] of Object.entries(pages)) {
+ for (const scale of [1, 3]) {
+ const file = join(work, `${name}.html`);
+ writeFileSync(file, `<!doctype html><meta charset="utf-8"><style>${css}</style><body>${page.html}</body>`);
+ const tab = await browser.newPage({ viewport: { width: page.w, height: page.h }, deviceScaleFactor: scale });
+ await tab.goto(pathToFileURL(file).href);
+ await tab.evaluate(() => document.fonts.ready);
+ await tab.waitForTimeout(200);
+ const png = join(work, `${name}@${scale}.png`);
+ await tab.screenshot({ path: png, clip: { x: 0, y: 0, width: page.w, height: page.h } });
+ await tab.close();
+ if (scale === 3 && args.preview) { mkdirSync(resolve(args.preview), { recursive: true }); writeFileSync(join(resolve(args.preview), `${name}.png`), readFileSync(png)); }
+ }
+}
+await browser.close();
+// 24-bit BMP, the format NSIS wants: render at 3x, shrink with a good filter
+const py = `import sys\nfrom PIL import Image\nfor name in ('sidebar','header'):\n im = Image.open(sys.argv[1] + '/' + name + '@3.png').convert('RGB')\n w, h = im.width // 3, im.height // 3\n im.resize((w, h), Image.LANCZOS).save(sys.argv[2] + '/' + name + '.bmp', 'BMP')\n`;
+const done = spawnSync('python', ['-c', py, work, out], { encoding: 'utf8' });
+if (done.status !== 0) { console.error(done.stderr); process.exit(1); }
+console.log('wrote', join(out, 'sidebar.bmp'), join(out, 'header.bmp'));
diff --git a/apps/desktop/scripts/ui-flows.mjs b/apps/desktop/scripts/ui-flows.mjs
index 3fc3f0e8f..69e261a8f 100644
--- a/apps/desktop/scripts/ui-flows.mjs
+++ b/apps/desktop/scripts/ui-flows.mjs
@@ -98,19 +98,23 @@ for line in sys.stdin:
`;
writeFileSync(join(control, 'tools', 'strata_mcp.py'), FAKE_INSTALLER);
-const FAKE_ENGINE = `import os, sys, json
+const FAKE_ENGINE = `import os, sys, json, time
from pathlib import Path
if os.environ.get('FAKE_ENGINE_FAIL'):
print('memory guard: 5.1 GiB of RAM free is under the 6 GiB limit (limit 63.1 GiB)', flush=True)
sys.exit(3)
-print('READY 4096 stop', flush=True)
+ctx = '4096'
+if '--max-context' in sys.argv: ctx = sys.argv[sys.argv.index('--max-context') + 1]
+print('READY %s stop' % ctx, flush=True)
for line in sys.stdin:
if line.strip() == 'QUIT': break
if line.startswith(('GEN ', 'BGEN ')):
slot = int(line.split()[1]) if line.startswith('BGEN ') else None
+ t0 = time.time(); n = 0
for ch in os.environ.get('FAKE_REPLY', 'a'):
- print('T %d' % ord(ch), flush=True)
- print('DONE 1 1 1.0 1.0 length 0 0 0', flush=True)
+ print('T %d' % ord(ch), flush=True); n += 1
+ if os.environ.get('FAKE_TOKEN_DELAY'): time.sleep(float(os.environ['FAKE_TOKEN_DELAY']))
+ print('DONE %d 12 300.0 %.1f length 0 0 0' % (n, max(1.0, (time.time() - t0) * 1000)), flush=True)
if slot is not None: print(f'BADM {slot} 0', flush=True)
`;
@@ -243,11 +247,22 @@ await scenario('models', { installed: true }, async ({ base, api, browser, shot
for (let i = 0; i < 90 && !ready; i++) { const s = (await api('/api/native')).body; ready = s.instances.some(x => x.status === 'ready'); const err = s.instances.find(x => x.status === 'error'); if (err) throw new Error('load failed: ' + err.error); await delay(1000); }
if (!ready) { const st = (await api('/api/native')).body; console.log('info jobs', JSON.stringify(st.jobs.map(j => [j.kind, j.state, j.error ?? j.summary])), 'instances', JSON.stringify(st.instances.map(i => [i.status, i.error]))); await shot(page, '2-models-after-click'); }
check(ready, 'the model loads (a fake engine under the real Control)');
+ await page.reload({ waitUntil: 'domcontentloaded' });
+ await page.getByRole('button', { name: /^Unload$/ }).first().waitFor({ timeout: 30000 }).catch(() => {});
+ await shot(page, '3-models-loaded');
+ check(await page.getByRole('button', { name: /^Unload$/ }).count() > 0, 'after a reload the Models screen shows the model as loaded, with Unload');
+ check(await page.getByRole('button', { name: /^Load model$/ }).count() === 0, 'and no longer offers Load model');
+ await page.getByRole('button', { name: /^Unload$/ }).first().click();
+ let gone = false;
+ for (let i = 0; i < 40 && !gone; i++) { gone = (await api('/api/native')).body.instances.length === 0; await delay(500); }
+ check(gone, 'Unload stops the engine');
+ await page.getByRole('button', { name: /^Load model$/ }).first().waitFor({ timeout: 15000 }).catch(() => {});
+ check(await page.getByRole('button', { name: /^Load model$/ }).count() > 0, 'and Load model is back');
check(page.problems.length === 0, 'no page error', page.problems.join('; '));
});
// ------------------------------------------------------------------------------------------------ tune + reading size
-await scenario('tune', { installed: true, env: { FAKE_REPLY: ['I am thinking about it.', '</think>', '', 'Here is a list:', '', '- one', '- two', '', '## Heading', '', 'Some `code` and **bold** text.', '', '| a | b |', '|---|---|', '| 1 | 2 |', '', '```', 'code line', '```', ''].join(String.fromCharCode(10)) } }, async ({ base, api, browser, dirs, shot }) => {
+await scenario('tune', { installed: true, env: { FAKE_TOKEN_DELAY: '0.02', FAKE_REPLY: ['I am thinking about it.', '</think>', '', 'Here is a list:', '', '- one', '- two', '', '## Heading', '', 'Some `code` and **bold** text.', '', '| a | b |', '|---|---|', '| 1 | 2 |', '', '```', 'code line', '```', ''].join(String.fromCharCode(10)) } }, async ({ base, api, browser, dirs, shot }) => {
const page = await newPage(browser, base);
await clickText(page, /get started/i);
await page.getByText(/Choose a version/i).waitFor({ timeout: 30000 });
@@ -272,23 +287,51 @@ await scenario('tune', { installed: true, env: { FAKE_REPLY: ['I am thinking abo
await page.locator('.st-msg--assistant .st-bubble').first().waitFor({ timeout: 60000 });
await page.locator('.st-msg--assistant .st-bubble p').first().waitFor({ timeout: 60000 }).catch(async () => console.log('info reply so far:', JSON.stringify(await page.locator('.st-msg--assistant').first().innerText()).slice(0, 300)));
await page.waitForFunction(() => !document.querySelector('.st-msg--assistant .caret'), null, { timeout: 90000 }).catch(() => {});
+ // the chat's telemetry chips: speed and context
+ const chips = page.locator('.telemetry__chip');
+ const chipTexts = await chips.allInnerTexts();
+ console.log('info chips', JSON.stringify(chipTexts));
+ let speedText = '', contextText = '';
+ for (let i = 0; i < await chips.count(); i++) {
+ await chips.nth(i).click();
+ await delay(400);
+ const pop = await page.locator('.telepop').last().innerText().catch(() => '');
+ if (/tok\/s|speed/i.test(pop) && !speedText) speedText = pop;
+ if (/context/i.test(pop) && !contextText) contextText = pop;
+ await page.keyboard.press('Escape');
+ }
+ const tps = Number(speedText.match(/([\d.]+)\s*tok(?:ens)?\/s/)?.[1] ?? NaN);
+ check(tps > 15 && tps < 120, 'the speed shown is the decode speed (the engine paced at ~50 tokens a second)', speedText.replace(/\s+/g, ' ').slice(0, 160));
+ check(/32\.8K|32K|32,768|32768/.test(contextText) && !/128K/.test(contextText), 'the context shown is the one the model runs with (32K), not 128K', contextText.replace(/\s+/g, ' ').slice(0, 160));
const sizes = [];
- for (const px of [12, 16, 22]) {
+ for (const px of [12, 14, 16, 22]) {
await page.evaluate(v => { const k = 'strata.prefs.v2'; const cur = JSON.parse(localStorage.getItem(k) || '{}'); localStorage.setItem(k, JSON.stringify({ ...cur, contentFontSize: v })); }, px);
await page.reload({ waitUntil: 'domcontentloaded' });
await page.locator('.st-msg--assistant .st-bubble').first().waitFor({ timeout: 30000 });
+ await page.screenshot({ path: join(shots, `tune-3-chat-${px}px.png`) });
sizes.push(await page.evaluate(() => { const b = document.querySelector('.st-msg--assistant .st-bubble'); const t = b.querySelector('p, .md-li, span') ?? b; const u = document.querySelector('.st-msg--user .st-bubble'); const all = {}; for (const el of b.querySelectorAll('*')) { if (!el.childNodes.length || ![...el.childNodes].some(n => n.nodeType === 3 && n.textContent.trim())) continue; all[el.tagName.toLowerCase() + (el.className ? '.' + String(el.className).split(' ')[0] : '')] = parseFloat(getComputedStyle(el).fontSize); } const extra = {}; for (const [k, q] of Object.entries({ reasoning: '.st-msg--assistant .st-collapse__body', codeblock: '.st-msg--assistant .st-code pre', codehead: '.st-msg--assistant .st-code__head' })) { const e = document.querySelector(q); if (e) extra[k] = parseFloat(getComputedStyle(e).fontSize); } Object.assign(all, extra); const chain = []; for (let e = b.querySelector('p'); e && e !== b; e = e.parentElement) chain.push(e.tagName + '.' + String(e.className).split(' ')[0] + '=' + getComputedStyle(e).fontSize); all.chain = chain; globalThis.__all = all; return [parseFloat(getComputedStyle(b).fontSize), parseFloat(getComputedStyle(t).fontSize), parseFloat(getComputedStyle(u).fontSize), all]; }));
}
console.log('info chain', JSON.stringify(await page.evaluate(() => { const el = document.querySelector('.st-msg--assistant .st-bubble p'); const out = []; for (let e = el; e && !e.classList.contains('st-msg'); e = e.parentElement) out.push(e.tagName + '.' + String(e.className).split(' ')[0] + '=' + getComputedStyle(e).fontSize); return out; })));
+ // which visible text does NOT follow the reading size? (same elements at 12 and 22)
+ const survey = async px => { await page.evaluate(v => { const k = 'strata.prefs.v2'; const cur = JSON.parse(localStorage.getItem(k) || '{}'); localStorage.setItem(k, JSON.stringify({ ...cur, contentFontSize: v })); }, px); await page.reload({ waitUntil: 'domcontentloaded' }); await page.locator('.st-msg--assistant .st-bubble').first().waitFor({ timeout: 30000 }); await delay(600);
+ return page.evaluate(() => { const out = {}; for (const el of document.querySelectorAll('main *, .chat *, textarea')) { if (![...el.childNodes].some(n => n.nodeType === 3 && n.textContent.trim()) && el.tagName !== 'TEXTAREA') continue; const r = el.getBoundingClientRect(); if (!r.width || !r.height) continue; const key = el.tagName.toLowerCase() + '.' + String(el.className).split(' ')[0] + ' <' + (el.parentElement?.className ?? '').toString().split(' ')[0] + '>'; out[key] = parseFloat(getComputedStyle(el).fontSize); } return out; }); };
+ const a = await survey(12), b = await survey(22);
+ // a fresh app (nothing saved) uses 14 px
+ await page.evaluate(() => localStorage.removeItem('strata.prefs.v2'));
+ await page.reload({ waitUntil: 'domcontentloaded' });
+ await page.locator('.st-msg--assistant .st-bubble').first().waitFor({ timeout: 30000 });
+ check(await page.evaluate(() => getComputedStyle(document.querySelector('.st-msg--assistant .st-bubble')).fontSize) === '14px', 'the default reading size is 14 px');
+ const fixed = Object.keys(a).filter(k => k in b && a[k] === b[k]);
+ console.log('info text that does not follow the reading size:', JSON.stringify(fixed.map(k => k + '=' + a[k])));
await shot(page, '2-chat-22px');
console.log('info sizes', JSON.stringify(sizes), JSON.stringify(await page.locator('.st-msg--assistant .st-bubble').first().innerHTML()).slice(0, 400));
- check(sizes[0][1] === 12 && sizes[1][1] === 16 && sizes[2][1] === 22 && sizes[0][2] === 12 && sizes[2][2] === 22, 'the reading size changes the text size of the conversation', JSON.stringify(sizes));
+ check(sizes[0][1] === 12 && sizes[1][1] === 14 && sizes[2][1] === 16 && sizes[3][1] === 22 && sizes[0][2] === 12 && sizes[3][2] === 22, 'the reading size changes the text size of the conversation', JSON.stringify(sizes));
// the slider itself
await page.evaluate(() => { location.hash = '#/chat?settings=1'; });
const slider = page.getByLabel('Reading text size');
if (await slider.count()) {
await slider.first().focus();
- for (let i = 0; i < 4; i++) await page.keyboard.press('ArrowLeft');
+ for (let i = 0; i < 4; i++) await page.keyboard.press('ArrowRight');
await delay(500);
const now = await page.evaluate(() => getComputedStyle(document.querySelector('.st-msg--assistant .st-bubble') ?? document.body).fontSize);
check(parseFloat(now) === 18, 'moving the slider changes the conversation text at once', now);
diff --git a/apps/desktop/src-tauri/icons/128x128.png b/apps/desktop/src-tauri/icons/128x128.png
index 63fb845f0..432eaf489 100644
Binary files a/apps/desktop/src-tauri/icons/128x128.png and b/apps/desktop/src-tauri/icons/128x128.png differ
diff --git a/apps/desktop/src-tauri/icons/[email protected] b/apps/desktop/src-tauri/icons/[email protected]
index 3417d3529..3bea2d8ff 100644
Binary files a/apps/desktop/src-tauri/icons/[email protected] and b/apps/desktop/src-tauri/icons/[email protected] differ
diff --git a/apps/desktop/src-tauri/icons/32x32.png b/apps/desktop/src-tauri/icons/32x32.png
index a3b6ecbad..c84b94653 100644
Binary files a/apps/desktop/src-tauri/icons/32x32.png and b/apps/desktop/src-tauri/icons/32x32.png differ
diff --git a/apps/desktop/src-tauri/icons/64x64.png b/apps/desktop/src-tauri/icons/64x64.png
index 6152d38ad..c8c67a042 100644
Binary files a/apps/desktop/src-tauri/icons/64x64.png and b/apps/desktop/src-tauri/icons/64x64.png differ
diff --git a/apps/desktop/src-tauri/icons/icon.ico b/apps/desktop/src-tauri/icons/icon.ico
index a052e27f4..e38a27ec3 100644
Binary files a/apps/desktop/src-tauri/icons/icon.ico and b/apps/desktop/src-tauri/icons/icon.ico differ
diff --git a/apps/desktop/src-tauri/icons/icon.png b/apps/desktop/src-tauri/icons/icon.png
index 3b9470607..da9c7603c 100644
Binary files a/apps/desktop/src-tauri/icons/icon.png and b/apps/desktop/src-tauri/icons/icon.png differ
diff --git a/apps/desktop/src-tauri/nsis/header.bmp b/apps/desktop/src-tauri/nsis/header.bmp
new file mode 100644
index 000000000..efc77717d
Binary files /dev/null and b/apps/desktop/src-tauri/nsis/header.bmp differ
diff --git a/apps/desktop/src-tauri/nsis/installer.nsi b/apps/desktop/src-tauri/nsis/installer.nsi
index cca93d5d9..0c6fe9aa0 100644
--- a/apps/desktop/src-tauri/nsis/installer.nsi
+++ b/apps/desktop/src-tauri/nsis/installer.nsi
@@ -165,6 +165,16 @@ VIAddVersionKey "ProductVersion" "${VERSION}"
!define MUI_LANGDLL_REGISTRY_KEY "${MANUPRODUCTKEY}"
!define MUI_LANGDLL_REGISTRY_VALUENAME "Installer Language"
+; Strata's look and words: the app's light surface and ink, short copy, no directory page, straight on to the progress.
+!define MUI_BGCOLOR "F9FAFB"
+!define MUI_TEXTCOLOR "171B1E"
+!define MUI_INSTFILESPAGE_PROGRESSBAR "smooth"
+!define MUI_WELCOMEPAGE_TITLE "State of the art. At home."
+!define MUI_WELCOMEPAGE_TEXT "Strata runs AI on your own computer.$\r$\n$\r$\nThis takes about a minute and needs no administrator. You choose your first model inside Strata."
+!define MUI_FINISHPAGE_TITLE "Strata is ready."
+!define MUI_FINISHPAGE_TEXT "Open it to pick your first model. Strata downloads it for you and gets it ready."
+!define MUI_FINISHPAGE_RUN_TEXT "Open Strata"
+
; Installer pages, must be ordered as they appear
; 1. Welcome Page
!define MUI_PAGE_CUSTOMFUNCTION_PRE SkipIfPassive
@@ -395,7 +405,7 @@ Function PageLeaveReinstall
FunctionEnd
; 5. Choose install directory page
-!define MUI_PAGE_CUSTOMFUNCTION_PRE SkipIfPassive
+!define MUI_PAGE_CUSTOMFUNCTION_PRE Skip
!insertmacro MUI_PAGE_DIRECTORY
; 6. Start menu shortcut page
@@ -415,7 +425,7 @@ Var AppStartMenuFolder
;
; Don't auto jump to finish page after installation page,
; because the installation page has useful info that can be used debug any issues with the installer.
-!define MUI_FINISHPAGE_NOAUTOCLOSE
+; (no NOAUTOCLOSE: welcome, progress, finish)
; Use show readme button in the finish page as a button create a desktop shortcut
!define MUI_FINISHPAGE_SHOWREADME
!define MUI_FINISHPAGE_SHOWREADME_TEXT "$(createDesktop)"
diff --git a/apps/desktop/src-tauri/nsis/sidebar.bmp b/apps/desktop/src-tauri/nsis/sidebar.bmp
new file mode 100644
index 000000000..9d936c6cf
Binary files /dev/null and b/apps/desktop/src-tauri/nsis/sidebar.bmp differ
diff --git a/apps/desktop/src-tauri/tauri.conf.json b/apps/desktop/src-tauri/tauri.conf.json
index 585c61cb1..daf8c6218 100644
--- a/apps/desktop/src-tauri/tauri.conf.json
+++ b/apps/desktop/src-tauri/tauri.conf.json
@@ -24,7 +24,8 @@
"deb"
],
"icon": [
- "icons/icon.png"
+ "icons/icon.png",
+ "icons/icon.ico"
],
"category": "DeveloperTool",
"shortDescription": "Local AI, tuned to one model.",
@@ -45,6 +46,9 @@
"nsis": {
"installMode": "currentUser",
"template": "nsis/installer.nsi",
+ "headerImage": "nsis/header.bmp",
+ "sidebarImage": "nsis/sidebar.bmp",
+ "installerIcon": "icons/icon.ico",
"compression": "lzma",
"languages": [
"English"
diff --git a/apps/host/server/inference.mjs b/apps/host/server/inference.mjs
index 1c2fbde14..c3735863a 100644
--- a/apps/host/server/inference.mjs
+++ b/apps/host/server/inference.mjs
@@ -278,6 +278,12 @@ export async function proxyCompletion({ baseUrl, message, history = [], systemPr
let streaming = true;
let writing = false;
let timings=null,finishReason=null;
+ /* The speed shown while it streams is the decode speed of everything generated, reasoning included, from the first token: tokens
+ (one per streamed chunk, or characters / 4 when the server batches) over the time since the first token, never the visible
+ answer over the whole wall time. The engine's own numbers replace it when the answer is done. */
+ let chunks=0,chars=0,lastStats=0;
+ const liveStats=()=>{if(supervised||firstTokenAt===null)return;const now=Date.now();if(now-lastStats<400)return;const seconds=(now-firstTokenAt)/1000;if(seconds<.5)return;lastStats=now;
+ const tokens=chars/Math.max(1,chunks)<8?chunks:Math.ceil(chars/4);onFrame({type:'stats',tokensOut:tokens,tps:Number((tokens/seconds).toFixed(1)),source:'stream'});};
const metricsTicker = supervised
? setInterval(() => {
@@ -325,12 +331,16 @@ export async function proxyCompletion({ baseUrl, message, history = [], systemPr
if (delta.reasoning_content) {
if (firstTokenAt === null) engine?.beginThinking();
if (firstTokenAt === null) firstTokenAt = Date.now();
+ chunks++; chars+=delta.reasoning_content.length;
onFrame({ type: 'reasoning', text: delta.reasoning_content });
+ liveStats();
}
if (delta.content) {
if (!writing) { engine?.beginGenerating(); writing = true; }
if (firstTokenAt === null) firstTokenAt = Date.now();
+ chunks++; chars+=delta.content.length;
onFrame({ type: 'delta', text: delta.content });
+ liveStats();
}
}
if (done) break;
diff --git a/apps/host/server/native-resources.mjs b/apps/host/server/native-resources.mjs
index 10f9dcc64..eda84f7f2 100644
--- a/apps/host/server/native-resources.mjs
+++ b/apps/host/server/native-resources.mjs
@@ -15,9 +15,12 @@ export function resourceSettings(config,hardware,profile,model={}){
args.push('--pool-workers',String(previous>0?Math.min(previous,workers):workers));
}
// Preserve calibrated worker counts and the installation's safety/vision reserve.
- const safety=Math.max(config.vision?2048:700,baseline.reserve);
const placement=Array.isArray(config.gpu)?config.gpu:[config.gpu??0];
const gpuMemory=(hardware?.gpus??[]).filter(g=>placement.includes(g.index)).map(g=>g.vram_gb).filter(n=>Number.isFinite(n)&&n>0);
+ /* The engine warns below ~1-1.5 GiB of free VRAM (the driver then pages GPU memory to system RAM and decoding falls to a few tokens a
+ second, #2065), so neither profile may reserve less: 1536 MiB on a card of 16 GB or less, 1024 above; with pictures 2048. */
+ const floor=gpuMemory.length&&Math.min(...gpuMemory)>16?1024:1536;
+ const safety=Math.max(config.vision?2048:floor,baseline.reserve);
const reserve=profile==='work'&&gpuMemory.length?Math.max(safety,Math.floor(Math.min(...gpuMemory)*1024*.2)):safety;
remove('--vram-reserve-mib');args.push('--vram-reserve-mib',String(reserve));
// RAM: "leave 20% for my work" keeps the model's RAM within 80% of the PC. A model whose experts do not fit there (the
diff --git a/apps/host/src/components/NativeModels.tsx b/apps/host/src/components/NativeModels.tsx
index f635d82bf..e728b9272 100644
--- a/apps/host/src/components/NativeModels.tsx
+++ b/apps/host/src/components/NativeModels.tsx
@@ -32,6 +32,8 @@ export function NativeModels({ onOtherModels, onReady, onImport, onConnectEngine
expertSection.current?.scrollIntoView({ block: 'start', behavior: 'smooth' }); }, [expert]);
const selectedFamily = families.find(f => f.family === family || f.sizes.some(size => size.sourceFamily === family)) ?? families[0], selectedSize = selectedFamily?.sizes.find(s => s.id === size) ?? selectedFamily?.sizes[0];
const installed = native?.installed?.models.find(m => m.model === selectedSize?.id);
+ // The model that is running is shown as running, with Unload: not offered for loading again.
+ const loadedSelected = native?.instances.find(i => i.model === selectedSize?.id && i.status === 'ready');
// A model that is already on this PC loads as it was prepared: the page shows its settings, so loading it never asks to download.
useEffect(() => { if (installed?.ready) {
if (installed.context)
@@ -101,7 +103,7 @@ export function NativeModels({ onOtherModels, onReady, onImport, onConnectEngine
{tr(active ? <NativeProgress job={active}/> : null)}
{tr(job?.state === 'error' ? <ErrorNotice message={job.error ?? job.summary} action={tr("Retry")} onRetry={() => void load()}/> : null)}
{tr(error ? <ErrorNotice message={error}/> : null)}
- <div className="native-model-actions">{tr(onReady && native?.instances.some(i => i.status === 'ready') ? <Btn variant="primary" onClick={onReady}>{tr("Start chatting")}</Btn> : <><Btn icon="gauge" disabled={!native?.instances.some(i => i.status === 'ready') || !!active} onClick={() => setBenchmark(true)}>{tr("Tune model")}</Btn><Btn variant="primary" icon="models" disabled={pending || !native?.available || !!active || !selectedSize || expert && deviceChoice === 'split' && gpus.length < 2} onClick={() => void load()}>{tr(pending ? 'Checking…' : installed?.ready ? 'Load model' : 'Download & load')}</Btn></>)}</div>
+ <div className="native-model-actions">{tr(loadedSelected ? <><Btn icon="gauge" disabled={!!active} onClick={() => setBenchmark(true)}>{tr("Tune model")}</Btn><Btn disabled={pending || !!active} onClick={() => void perform(() => nativeAction(`instances/${loadedSelected.id}/unload`))}>{tr("Unload")}</Btn>{tr(onReady ? <Btn variant="primary" onClick={onReady}>{tr("Start chatting")}</Btn> : <span className="native-ready">{tr("Ready")}</span>)}</> : onReady && native?.instances.some(i => i.status === 'ready') ? <Btn variant="primary" onClick={onReady}>{tr("Start chatting")}</Btn> : <><Btn icon="gauge" disabled={!native?.instances.some(i => i.status === 'ready') || !!active} onClick={() => setBenchmark(true)}>{tr("Tune model")}</Btn><Btn variant="primary" icon="models" disabled={pending || !native?.available || !!active || !selectedSize || expert && deviceChoice === 'split' && gpus.length < 2} onClick={() => void load()}>{tr(pending ? 'Checking…' : installed?.ready ? 'Load model' : 'Download & load')}</Btn></>)}</div>
</div>
<aside className="stack stack--lg"><section className="st-card"><h2 className="h4">{tr("This computer")}</h2><p>{tr(native?.hardware?.ram_gb || hardware?.memory.totalGb || '—')}{tr(" GB RAM")}</p>{tr(native?.hardware?.gpus.length ? native.hardware.gpus.map(g => <p className="small" key={g.index}>{tr(g.name)} · {tr(g.vram_gb)}{tr(" GB")}</p>) : <p className="small muted">{tr("No supported GPU reported by this installation.")}</p>)}<p className="small muted">{tr("Compatibility and sizes come from your installed Strata version.")}</p></section><section className="st-card"><h2 className="h4">{tr("Model files")}</h2><p className="small muted">{tr("Strata prepares all shards, the draft layer and shared files. Completed downloads are reused.")}</p><code className="native-path">{installed?.dataFolder ?? native?.downloadDirectory ?? native?.installed?.data_folder ?? 'Choose an installation'}</code><Btn size="sm" variant="ghost" onClick={() => { setSettingsSection('Storage'); go('settings'); }}>{tr("Storage settings")}</Btn></section><Btn size="sm" variant="ghost" onClick={onOtherModels}>{tr("Other local models\u2026")}</Btn></aside>
</div>
diff --git a/apps/host/src/components/TelemetryPopover.tsx b/apps/host/src/components/TelemetryPopover.tsx
index 15522f8a7..f44e1ccf4 100644
--- a/apps/host/src/components/TelemetryPopover.tsx
+++ b/apps/host/src/components/TelemetryPopover.tsx
@@ -6,6 +6,7 @@ import { ContextRing } from './ContextRing';
import { Popover } from './Popover';
import { chatMetrics, compactTokens, elapsed } from '../lib/chat-metrics';
import { useStore } from '../lib/store';
+import { useNative } from '../lib/native';
import type { Session } from '../lib/types';
function Chip({ icon, visual, label, title, children }: {
icon?: IconName;
@@ -46,7 +47,11 @@ export function ChatTelemetry({ session }: {
}) {
const monitor = useStore(s => s.monitor);
const engine = useStore(s => s.engine);
- const m = chatMetrics(session, monitor, engine ? contextTokens(engine.beginner.contextLength) : undefined);
+ // The context shown is the one the loaded model runs with (the engine's own --max-context), not a catalog or settings default.
+ const native = useNative();
+ const running = native?.instances.find(i => i.id === native.defaultId && i.status === 'ready') ?? native?.instances.find(i => i.status === 'ready');
+ const runningContext = running ? Number(running.metrics?.engine?.max_context) || running.context || undefined : undefined;
+ const m = chatMetrics(session, monitor, runningContext ?? (engine ? contextTokens(engine.beginner.contextLength) : undefined));
const speed = useCalmSpeed(m.speed, session?.id ?? null, m.streaming);
if (!m.hasReplies)
return null;
diff --git a/apps/host/src/lib/store.ts b/apps/host/src/lib/store.ts
index 18703366e..f2ae03074 100644
--- a/apps/host/src/lib/store.ts
+++ b/apps/host/src/lib/store.ts
@@ -141,7 +141,7 @@ const PREFS_DEFAULTS: Prefs = {
contrast: 'standard',
rail: false,
focus: false,
- contentFontSize: 16,
+ contentFontSize: 14,
density: 'comfortable',
motion: 'system',
transparency: 'translucent',
@@ -188,7 +188,7 @@ function loadPrefs(): Prefs {
...parsed,
language: validLanguage(parsed.language) ? parsed.language : 'en',
fontFamily: typeof parsed.fontFamily === 'string' ? parsed.fontFamily.slice(0,128).replace(/[\x00-\x1f]/g,'') : '',
- contentFontSize: Math.min(CONTENT_FS_MAX, Math.max(CONTENT_FS_MIN, Number(parsed.contentFontSize) || 16)),
+ contentFontSize: Math.min(CONTENT_FS_MAX, Math.max(CONTENT_FS_MIN, Number(parsed.contentFontSize) || 14)),
};
} catch {
/* fall through to the next key */
@@ -1353,7 +1353,7 @@ export async function sendMessage(text: string) {
} else if (type === 'agent') {
patchSession(sessionId, s => ({ ...s, agents: updateAgent(s.agents ?? [], frame.agent) }));
} else if (type === 'delta') {
- patchAssistant((m) => { const content=m.content+String(frame.text??''); const count=Math.max(1,Math.ceil(content.length/4)); const seconds=Math.max(.1,(performance.now()-started)/1000); return {...m,content,tokens:count,inputTokens:m.inputTokens??Math.ceil((body.length+prior.reduce((n,p)=>n+p.content.length,0))/4),tps:count/seconds,usageSource:'estimated'}; });
+ patchAssistant((m) => { const content=m.content+String(frame.text??''); const count=Math.max(1,Math.ceil((content.length+(m.reasoning?.length??0))/4)); return {...m,content,tokens:m.usageSource==='estimated'&&m.tps?m.tokens:count,inputTokens:m.inputTokens??Math.ceil((body.length+prior.reduce((n,p)=>n+p.content.length,0))/4),usageSource:'estimated'}; });
} else if (type === 'artifacts' && Array.isArray(frame.artifacts)) {
patchAssistant(m => ({ ...m, artifacts: frame.artifacts as PreviewArtifact[] }));
} else if(type==='native-tool'){
diff --git a/apps/host/src/styles/app.css b/apps/host/src/styles/app.css
index b37f74af0..0c7c1f2b7 100644
--- a/apps/host/src/styles/app.css
+++ b/apps/host/src/styles/app.css
@@ -452,11 +452,11 @@
the text size — a fixed px allowance drifts at the extremes. */
.app[data-focus='true'] .slashmenu,
.app[data-focus='true'] .chat__inner {
- font-size: var(--st-content-fs, 16px);
+ font-size: var(--st-content-fs, 14px);
max-width: calc(var(--st-measure) + 2.5em);
}
.app[data-focus='true'] .composer__inner {
- font-size: var(--st-content-fs, 16px);
+ font-size: var(--st-content-fs, 14px);
/* the same column, less the transcript's 28px padding on each side */
max-width: calc(var(--st-measure) + 2.5em - 56px);
}
@@ -474,8 +474,8 @@
— the same scoping DSH uses for its 10–22px content setting. */
.st-msg--assistant .st-bubble,
.st-msg--user .st-bubble {
- font-size: var(--st-content-fs, 16px);
- line-height: calc(var(--st-content-fs, 16px) * 1.62);
+ font-size: var(--st-content-fs, 14px);
+ line-height: calc(var(--st-content-fs, 14px) * 1.62);
}
/* Density — the difference between scanning data and reading it. */
@@ -6998,10 +6998,13 @@ textarea {
model's reasoning keep their usual proportion (18px, 13px, 13px at the 16px default) at any size. */
.st-bubble h2, .st-bubble h3, .st-bubble h4 { font-size: 1.125em; }
.st-bubble .md-code { font-size: .8125em; }
-.st-bubble .md-table { font-size: calc(var(--st-content-fs, 16px) * .8125); }
+.st-bubble .md-table { font-size: calc(var(--st-content-fs, 14px) * .8125); }
.st-bubble .md-table th, .st-bubble .md-table td { font-size: inherit; }
-.st-bubble .st-code pre, .st-bubble .st-code pre code, .st-bubble .st-code .syntax-text { font-size: calc(var(--st-content-fs, 16px) * .8125); }
-.st-msg--assistant .st-collapse__body { font-size: calc(var(--st-content-fs, 16px) * .8125); }
+.st-bubble .st-code pre, .st-bubble .st-code pre code, .st-bubble .st-code .syntax-text { font-size: calc(var(--st-content-fs, 14px) * .8125); }
+.st-msg--assistant .st-collapse__body { font-size: calc(var(--st-content-fs, 14px) * .8125); }
.error-notice .error-detail { margin-top: 6px; }
.error-notice .error-detail pre { white-space: pre-wrap; overflow-wrap: anywhere; margin: 6px 0 0; font-family: var(--st-font-mono); }
.error-detail-text { white-space: pre-wrap; overflow-wrap: anywhere; margin: 6px 0 0; font-family: var(--st-font-mono); }
+/* What you type and the small print under each message follow the reading size too, so the whole conversation changes together. */
+.st-composer textarea { font-size: var(--st-content-fs, 14px); }
+.st-msg__meta, .st-msg__meta time { font-size: calc(var(--st-content-fs, 14px) * .8125); }
diff --git a/apps/host/tests/agent-host.test.mjs b/apps/host/tests/agent-host.test.mjs
index 7335f89fd..1fabf6995 100644
--- a/apps/host/tests/agent-host.test.mjs
+++ b/apps/host/tests/agent-host.test.mjs
@@ -26,7 +26,7 @@ const sleep = ms => new Promise(r => setTimeout(r, ms));
async function startHost(t, { home, ws, harness = HARNESS } = {}) {
const port = await freePort();
- const env = { ...process.env, STRATA_HOME: home, STRATA_PORT: String(port), STRATA_WORKSPACE: ws, STRATA_REAL_ENGINE: '0' };
+ const env = { ...process.env, STRATA_HOME: home, STRATA_PORT: String(port), STRATA_WORKSPACE: ws, STRATA_REAL_ENGINE: '0', STRATA_COMPONENTS_DIR: join(home, 'components') };
if (harness) env.STRATA_HARNESS_DIR = harness; else delete env.STRATA_HARNESS_DIR;
const child = spawn(process.execPath, [HOST], { env, stdio: ['ignore', 'pipe', 'pipe'], windowsHide: true });
let log = '';
diff --git a/apps/host/tests/resource-ram.test.mjs b/apps/host/tests/resource-ram.test.mjs
index bac52b5de..3965ea479 100644
--- a/apps/host/tests/resource-ram.test.mjs
+++ b/apps/host/tests/resource-ram.test.mjs
@@ -29,3 +29,17 @@ test('a model that fits, or one that already has a RAM mode, is left alone', ()
assert.equal(resourceSettings(own, pc64, 'work', iq3s).args.includes('--resident-budget-gib'), false);
assert.equal(resourceSettings(config, pc64, 'work').args.includes('--resident-budget-gib'), false, 'no model size, no change');
});
+
+test('neither profile reserves less VRAM than the engine\'s safe floor (the driver pages GPU memory below it and decoding collapses)', () => {
+ const card12 = { cpu: { cores: 12 }, ram_gb: 63, gpus: [{ index: 0, vram_gb: 11.9 }] };
+ const card24 = { cpu: { cores: 12 }, ram_gb: 63, gpus: [{ index: 0, vram_gb: 24 }] };
+ for (const profile of ['work', 'performance']) {
+ const lowConfig = { gpu: 0, args: ['--max-context', '32768', '--vram-reserve-mib', '700'] };
+ assert.ok(resourceSettings(lowConfig, card12, profile).vram_reserve_mib >= 1536, `${profile} on 12 GB`);
+ assert.ok(resourceSettings({ gpu: 0, args: [] }, card12, profile).vram_reserve_mib >= 1536, `${profile}, nothing configured`);
+ assert.ok(resourceSettings(lowConfig, card24, profile).vram_reserve_mib >= 1024, `${profile} on 24 GB`);
+ }
+ assert.equal(resourceSettings({ gpu: 0, args: [] }, card12, 'performance').vram_reserve_mib, 1536);
+ assert.equal(resourceSettings({ gpu: 0, args: [] }, card24, 'performance').vram_reserve_mib, 1024);
+ assert.equal(resourceSettings({ gpu: 0, args: [], vram_reserve_mib: 2437 }, card12, 'performance').vram_reserve_mib, 2437, 'a larger reserve of the installation is kept');
+});
diff --git a/apps/host/tests/stream-speed.test.mjs b/apps/host/tests/stream-speed.test.mjs
new file mode 100644
index 000000000..309fedc76
--- /dev/null
+++ b/apps/host/tests/stream-speed.test.mjs
@@ -0,0 +1,40 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { createServer } from 'node:http';
+import { proxyCompletion } from '../server/inference.mjs';
+
+/* The speed shown while an answer streams must be the decode speed of everything generated, reasoning included, from the first
+ token (the owner saw 4-5 tok/s in chat while the engine log said 48-57): the stand-in engine streams 90 reasoning tokens at about
+ 40 a second, a 4 s "prompt" before the first one, then a short answer. */
+function engine({ promptMs, reasoning, answer, perSecond }) {
+ return createServer((req, res) => {
+ res.writeHead(200, { 'content-type': 'text/event-stream' });
+ const send = obj => res.write(`data: ${JSON.stringify(obj)}\n\n`);
+ let sent = 0;
+ setTimeout(function tick() {
+ if (sent < reasoning + answer) {
+ send({ choices: [{ delta: sent < reasoning ? { reasoning_content: 'tok ' } : { content: 'ans ' } }] });
+ sent++;
+ return setTimeout(tick, 1000 / perSecond);
+ }
+ send({ choices: [], usage: { prompt_tokens: 20, completion_tokens: reasoning + answer } });
+ res.write('data: [DONE]\n\n'); res.end();
+ }, promptMs);
+ });
+}
+
+test('live speed counts reasoning tokens over decode time, not the answer over the whole wait', async () => {
+ const server = engine({ promptMs: 4000, reasoning: 90, answer: 10, perSecond: 40 });
+ await new Promise(r => server.listen(0, '127.0.0.1', r));
+ const frames = [];
+ try {
+ const result = await proxyCompletion({ baseUrl: `http://127.0.0.1:${server.address().port}/v1`, message: 'hi', onFrame: f => frames.push(f), engine: { beginReading() {}, beginThinking() {}, beginGenerating() {} } });
+ const stats = frames.filter(f => f.type === 'stats');
+ assert.ok(stats.length >= 2, 'live frames while it streams');
+ const last = stats.at(-1);
+ assert.ok(last.tps > 22 && last.tps < 80, `live tok/s ${last.tps} is the decode rate, not tokens over the wait including the prompt (~14)`);
+ assert.ok(last.tokensOut >= 50, 'reasoning tokens are counted');
+ assert.equal(result.tokensOut, 100);
+ assert.ok(result.tps > 22 && result.tps < 80, `final tok/s ${result.tps}`);
+ } finally { server.close(); }
+});