[Bug]: Remote preview stays on "Connecting to browser…" when the viewer's first still capture never answers #24852

Activity

1. xIGBClutchIx commented on Oct 25, 2026

   

   xIGBClutchIx

   on Oct 25, 2026

   More actions

   Glad to hear I am not the only one having issues. Just started using T3 Code. It also really wanted to open the dev url as a file.
2. ryanskybluechem commented on Nov 5, 2026

   

   ryanskybluechem

   on Nov 5, 2026

   More actions

   Still reproducing on 0.1.8-nightly.20261105.4218 (macOS arm64, desktop app host, viewer via T3 Connect relay). Today's source still has the unbounded pushStill before the upgrade (ServerBrowser.ts:2683, captureViewport in ServerBrowserPage.ts:469 has no deadline).

   Host server.trace.ndjson, 2026-11-05 UTC:

   15:20–15:30  tab_4  ServerBrowser.attachViewer  35–53 ms      ok
   15:31:19     tab_5  ServerBrowser.attachViewer  55,474 ms     GET /api/preview-stream/ws interrupted
   15:31:31     tab_5  ServerBrowser.attachViewer  43,539 ms     interrupted
   15:31:40     tab_5  ServerBrowser.attachViewer  34,270 ms     interrupted
   

   The relay connector logs Incoming request ended abruptly: context canceled for each preview-stream request at the same moments (15:31:31, 15:31:34, 15:37:51, 15:38:42), so the viewer gives up while the server is still waiting on the still. A freshly opened tab works until its renderer stops answering. Open fixes: #27663, #25381, #25862.

   Triaged with Claude Opus 5.5 in T3 Code.
3. alexispurslane commented on Nov 9, 2026

   

   alexispurslane

   on Nov 9, 2026

   More actions

   I reproduced the trigger on the headless server browser. A viewer's still capture that is pending when the tab commits a navigation to a different site never gets an answer. pushStill then holds tab.captureLock until the tab closes, and the tab wedges as this issue describes. This answers "Why the first renderer stopped answering is not reproduced" above. The renderer did not stop. Chromium dropped the request when it swapped renderer processes.

   Standalone reproduction

   The script below does not use T3. It uses playwright-core 1.60.0 and the chrome-headless-shell 156.0.8078.70 that T3 downloads. A second CDP session plays the viewer. It starts Page.captureScreenshot while page script is busy, and then the page navigates. 127.0.0.1 and localhost are different sites.

   Navigation while the capture is pending	Pending capture	New capture on the new page
   Same site	answered	answered in 35 ms
   Different site	no answer after 30 s (180 s in a longer run)	answered in 12 ms
   Different site, then CDPSession.detach()	rejected at the detach	answered in 8 ms

   On the same busy page, Page.stopScreencast and Page.startScreencast answered in under 5 ms, so the viewer's pause and resume are not the hang. Page.captureScreenshot, Runtime.evaluate and key presses waited until the page script finished.

   node lost-capture-repro.cjs <playwright-core dir> <chrome-headless-shell path>

   // node lost-capture-repro.cjs <playwright-core dir> <chrome-headless-shell path>
   // A viewer-style CDP session starts Page.captureScreenshot while the page is busy, then the
   // page navigates. 127.0.0.1 and localhost are different sites, so Chromium swaps processes.
   const http = require("node:http");
   const { chromium } = require(process.argv[2]);
   const busy = (ms) => `<script>setTimeout(() => { const e = Date.now() + ${ms}; while (Date.now() < e); }, 200)</script>`;
   const server = http.createServer((req, res) => {
     res.setHeader("content-type", "text/html");
     res.end(req.url === "/busy" ? `<h1>busy</h1>${busy(5000)}` : "<h1>idle</h1>");
   });
   const settle = (p, ms) => {
     const t0 = Date.now();
     return Promise.race([
       p.then(() => `answered after ${Date.now() - t0} ms`, (e) => `rejected after ${Date.now() - t0} ms: ${e.message.split("\n")[0]}`),
       new Promise((r) => setTimeout(() => r(`no answer after ${ms} ms`), ms)),
     ]);
   };
   (async () => {
     await new Promise((r) => server.listen(0, "127.0.0.1", r));
     const port = server.address().port;
     const browser = await chromium.launch({ executablePath: process.argv[3] });
     const context = await browser.newContext();
     for (const [label, target, detach] of [
       ["same site", `http://127.0.0.1:${port}/idle`, false],
       ["other site", `http://localhost:${port}/idle`, false],
       ["other site, then detach", `http://localhost:${port}/idle`, true],
     ]) {
       const page = await context.newPage();
       await page.goto(`http://127.0.0.1:${port}/busy`);
       const viewer = await context.newCDPSession(page);
       await new Promise((r) => setTimeout(r, 600));
       const still = viewer.send("Page.captureScreenshot", { format: "jpeg" });
       await new Promise((r) => setTimeout(r, 500));
       await page.goto(target);
       if (detach) setTimeout(() => viewer.detach().catch(() => {}), 8000);
       console.log(`${label.padEnd(24)} pending capture: ${await settle(still, 30000)}`);
       const fresh = await context.newCDPSession(page);
       console.log(`${"".padEnd(24)} new capture on new page: ${await settle(fresh.send("Page.captureScreenshot"), 5000)}`);
       await page.close();
     }
     await browser.close();
     server.close();
   })();

   How it wedges the tab (main @ 4f0b74e)

   1. pushStill (ServerBrowser.ts:2625) awaits captureViewport (ServerBrowserPage.ts:469) inside withCaptureLock. The capture has no deadline.
   2. The capture waits while page script is busy. On a heavy page, this window is several seconds wide.
   3. A navigation to a different site in that window loses the capture. The agent's preview_navigate, a click on an off-site link, or a redirect can cause it.
   4. The next preview_snapshot waits for the lock inside SessionControl.agent, before its 10 s bound starts. It never answers.
   5. Every later action on the tab queues behind that snapshot. The deadline check runs only when an entry starts, so these entries never start. Each one costs its full timeout plus 1 s, and a host eviction. Only preview_status, preview_dialog and t3_preview_close still work.

   pushStill runs when a viewer attaches before its first frame. It also runs 200 ms after screencast motion or a dropped frame. So the trigger needs a viewer, a busy page and a navigation to another site at the same time. This explains why the failure is intermittent, and why it hits one thread's tab while other tabs keep working.

   Field evidence

   Host: desktop app 0.1.6-nightly.20261031.3977 with the headless server browser, macOS 27.0 arm64, Codex agent. Provider log, UTC start times:

   2026-11-09, tab_11, one thread
   13:32:47  preview_navigate       967 ms  ok, visible:false
   13:33:04  preview_snapshot     8,026 ms  "The page was too busy to capture a snapshot in time" (bot-check page)
   13:33:18  preview_evaluate        23 ms  ok
   13:33:28  preview_navigate       832 ms  ok, to a different site, visible:true (a viewer attached since the last navigate)
   13:33:34  preview_snapshot    16,003 ms  timed out, host evicted
   13:33:54  preview_evaluate    16,009 ms  timed out, host evicted
   13:34:57  preview_navigate    21,010 ms  timed out, host evicted
             tab_12 in another thread worked normally on the same host the whole time
   

   Two more wedges on 2026-11-06 follow the same pattern: a snapshot with no answer, then evaluate and navigate with no answer, preview_status instant, and a new tab that works at once. One of them had visible: true. The logs for those two do not show the navigation.

   For the open fixes

   The lost request never settles by itself. A bound on captureViewport, as in #23515 and #25862, covers this trigger. Without a bound, only a session detach or a tab close frees the lock. The viewer's finalizer detaches its session, so a viewer that disconnects also frees a wedged tab. I did not test that path in the app.

   Investigated with Claude Opus 5.5 in T3 Code.
4. hunterbohm commented on Nov 10, 2026

   

   hunterbohm

   on Nov 10, 2026

   More actions

   Note

   Posted by Claude Code (Fable 5.2) on behalf of the account owner.

   Still reproduces on desktop Nightly 0.1.9-nightly.20261110.4478 (macOS 26.6.2, desktop app host, iOS viewer over T3 Connect), with a trigger that links this to #23139.

   • An agent opened a tab while the host window was off screen, and a phone attached as a viewer. The phone stayed on "Connecting to browser…".
   • The tab stayed stuck for over two hours. That included time when the host window was on screen and the tab was showing in the desktop browser panel.
   • During that time, Page.captureScreenshot sent straight to the same guest over the DevTools port answered in 26 ms. So the renderer was healthy and only the tab's viewer path was wedged.
   • A second tab opened in the same thread streamed to the phone normally.

   So one capture that hangs because of #23139 leaves the tab unusable for viewers until it is closed, even after the window is visible again.
5. juliusmarminge

   closed this as completedin #25381on Nov 13, 2026