Pup - the AI's own browser

@@ -15,11 +15,38 @@ const chrome = require('./chrome'); // Chrome-for-Testing self-heal: detect / au // v1.8.40+: previously this was wrapped in try/catch with a `let sharp = null` // fallback. The fallback path silently sent 2559px PNGs to /tmp/adom-desktop-screenshots/, // which CRASHES Claude many-image sessions at the 2000px limit. Hard-require-// sharp now: if it can't load, the bridge fails fast at startup with a clear-// error rather than producing oversized screenshots indefinitely.-const sharp = require('sharp');+// sharp is OPTIONAL — a heavy native dep whose install can fail on a fresh box (no+// prebuild / no build tools / a managed-runtime PATH gap). pup must still spawn and+// work without it: we fall back to serving the screenshot at capture size and reading+// dimensions from the PNG header. When sharp IS present, we downscale to keep images+// under Claude's Read limit (the ideal path). Never let a missing sharp crash the bridge.+let sharp = null;+try { sharp = require('sharp'); }+catch (e) { console.warn('[pup] sharp unavailable — screenshots served at capture size, not downscaled:', e.message); } const MAX_SCREENSHOT_DIM = 1568; // Claude's optimal max dimension +// PNG width/height straight from the IHDR header (bytes 16-23, big-endian). Pure JS,+// no sharp needed. puppeteer screenshots are always PNG.+function pngDimensions(buf) {+  try { return { width: buf.readUInt32BE(16), height: buf.readUInt32BE(20) }; }+  catch { return { width: 0, height: 0 }; }+}+// Downscale a PNG to <= maxDim (long edge) when sharp is available; otherwise pass it+// through at capture size. Returns the (possibly-resized) buffer + dimensions + flags.+async function processScreenshot(buf, maxDim) {+  if (sharp) {+    const orig = await sharp(buf).metadata();+    const needsResize = orig.width > maxDim || orig.height > maxDim;+    const out = await (needsResize+      ? sharp(buf).resize({ width: maxDim, height: maxDim, fit: 'inside' })+      : sharp(buf)).png({ compressionLevel: 9, adaptiveFiltering: true }).toBuffer();+    const fin = await sharp(out).metadata();+    return { buf: out, width: fin.width, height: fin.height, origWidth: orig.width, origHeight: orig.height, resized: needsResize, sharpUsed: true };+  }+  const d = pngDimensions(buf);+  return { buf, width: d.width, height: d.height, origWidth: d.width, origHeight: d.height, resized: false, sharpUsed: false };+}+ const PORT = (() => {   const idx = process.argv.indexOf('--port');   return idx !== -1 && process.argv[idx + 1] ? parseInt(process.argv[idx + 1]) : 8851;@@ -2827,22 +2854,12 @@ const server = http.createServer(async (req, res) => {             20_000, 'page.screenshot'           ); -          // v1.8.40+: sharp is mandatory — no `if (sharp)` guard, no try/catch-          // fallback. If the resize fails the request errors out cleanly rather-          // than silently sending a 2559px image that would crash a many-image-          // Claude session at the 2000px API limit. For the rare "I want the-          // real resolution to save to disk" case, use browser_screenshot_full_res.-          const origMeta = await sharp(screenshotBuf).metadata();+          // Downscale to keep the image under Claude's Read limit when sharp is+          // available; otherwise serve at capture size (sharp is optional — see the+          // processScreenshot helper). A ≤1568px image is the ideal path.           const maxDim = Math.min(maxWidth, MAX_SCREENSHOT_DIM);-          const needsResize = origMeta.width > maxDim || origMeta.height > maxDim;-          const pipeline = needsResize-            ? sharp(screenshotBuf).resize({ width: maxDim, height: maxDim, fit: 'inside' })-            : sharp(screenshotBuf);-          screenshotBuf = await pipeline.png({ compressionLevel: 9, adaptiveFiltering: true }).toBuffer();-          // v1.8.40+: probe the FINAL dimensions (post-resize) so the response-          // tells the caller exactly what they're going to Read. Without these-          // the AI can only guess from sizeKB.-          const finalMeta = await sharp(screenshotBuf).metadata();+          const shot = await processScreenshot(screenshotBuf, maxDim);+          screenshotBuf = shot.buf;            const filePath = path.join(SHOTS_DIR, `pup-${ts}.png`);           fs.writeFileSync(filePath, screenshotBuf);@@ -2855,11 +2872,12 @@ const server = http.createServer(async (req, res) => {             sizeKB,             filePath,             base64,-            width: finalMeta.width,-            height: finalMeta.height,-            origWidth: origMeta.width,-            origHeight: origMeta.height,-            resized: needsResize,+            width: shot.width,+            height: shot.height,+            origWidth: shot.origWidth,+            origHeight: shot.origHeight,+            resized: shot.resized,+            ...(shot.sharpUsed ? {} : { sharpUnavailable: true, _hint: 'Screenshot served at CAPTURE size — the sharp resize library is not installed on this box, so the image was NOT downscaled and may exceed Claude\'s ~2000px Read limit. If a Read fails on size, capture a smaller viewport, or (rare) this box\'s native-dep install failed — see browser_readiness.' }),             sessionId: resolvedId,             tabId,           }) });@@ -2885,10 +2903,9 @@ const server = http.createServer(async (req, res) => {             tab.page.screenshot({ type: 'png', fullPage }),             20_000, 'page.screenshot'           );-          // Probe dimensions for the warning payload — cheap (sharp just reads-          // the PNG header). No resize, no recompression — the bytes that hit-          // disk are exactly what Chrome captured.-          const meta = await sharp(screenshotBuf).metadata();+          // Probe dimensions for the warning payload — cheap. No resize, no+          // recompression — the bytes that hit disk are exactly what Chrome captured.+          const meta = sharp ? await sharp(screenshotBuf).metadata() : pngDimensions(screenshotBuf);           const filePath = path.join(SHOTS_DIR, `pup-fullres-${ts}.png`);           fs.writeFileSync(filePath, screenshotBuf);           const sizeKB = Math.round(screenshotBuf.length / 1024);@@ -3886,15 +3903,9 @@ const server = http.createServer(async (req, res) => {       const ts = Date.now();       const fp = url.searchParams.get('fullPage') === 'true';       let shotBuf = await session.page.screenshot({ type: 'png', fullPage: fp });-      // v1.8.40+: resize is mandatory (no guard, no fallback). See block at-      // top of file + browser_screenshot case for context.-      const origMeta = await sharp(shotBuf).metadata();-      const needsResize = origMeta.width > MAX_SCREENSHOT_DIM || origMeta.height > MAX_SCREENSHOT_DIM;-      const pipeline = needsResize-        ? sharp(shotBuf).resize({ width: MAX_SCREENSHOT_DIM, height: MAX_SCREENSHOT_DIM, fit: 'inside' })-        : sharp(shotBuf);-      shotBuf = await pipeline.png({ compressionLevel: 9, adaptiveFiltering: true }).toBuffer();-      const finalMeta = await sharp(shotBuf).metadata();+      // Downscale when sharp is available; else serve at capture size (sharp optional).+      const shot = await processScreenshot(shotBuf, MAX_SCREENSHOT_DIM);+      shotBuf = shot.buf;       const filePath = path.join(SHOTS_DIR, `pup-${ts}.png`);       fs.writeFileSync(filePath, shotBuf);       const sizeKB = Math.round(shotBuf.length / 1024);@@ -3921,11 +3932,11 @@ const server = http.createServer(async (req, res) => {         ok: true,         sizeKB,         filePath,-        width: finalMeta.width,-        height: finalMeta.height,-        origWidth: origMeta.width,-        origHeight: origMeta.height,-        resized: needsResize,+        width: shot.width,+        height: shot.height,+        origWidth: shot.origWidth,+        origHeight: shot.origHeight,+        resized: shot.resized,         container,       });       return;