The browser subagent cannot use more than one tab, so a five-seat check ran for an hour #118
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
The same run as #116 (run_01M3WX1VW43K77HWNQ328RBGJK, agent.test.browser). The app under test is a five-player card game, so the check needs five players on screen at once. The
browsersubagent's tool list leaves out Playwright'stabstool (13 of the server's 24 tools are allowed, trimmed in #17 for context cost), so the subagent drove five seats from one tab withwindow.openhandles andevaluate, built DOM copies of the other tabs to screenshot, hit a stray seat in the room, and spent two runs of about 22 minutes each, 43 of the step's 60 minutes, on it.Asked for:
browsersubagent thetabstool, and tell it to open one tab per player or window instead of reaching into other pages from one.Shipped and deployed, and archived as
browser-tabs. Commita3182a9, live sinceac2caaf(2026-10-02 05:12 UTC); production now runs3bc80ce. The limit is doubled as asked.browsersubagent has Playwright'stabstool (list, open, switch, close). It costs 630 more characters of the subagent's tools (the allowed set is 9,881), and its instructions say to open one tab for each person a check needs and to switch between them, and not to reach into one page from another. The primary's prompt says to ask for one tab per person when a check needs several people at once. Tabs share one browser context, so they do not give two people separate cookies.agent.test.browsergoes from 3,600 s to 7,200 s in every built-in pipeline. A run keeps the pipeline it started with (#102), so only runs admitted from now on have it; the engine's retry of a time-out (#119) doubles the limit either way.Seen on production: the resumed run's regenerated opencode.json lists the tool for the subagent. Not seen: the tool in use. The resumed attempt's primary was shown the previous subagent's note that its tools were pinned to one tab (true until this shipped) in the briefing of its resume, and wrote the
window.openworkaround into the new subagent's task. The briefing now ends its quoted reports with a line saying they describe what their writers had to work with then, and the browser check's step 3 tells the primary never to say the tools are pinned to one tab (91ad99f, deployed in3bc80ce). Task 3.4 stays open in the archived change until a production browser check calls the tool. If the next one still opens its seats withwindow.open, reopen this.