[modacker]

Change

Two budgets were too small for the machine the browser runs on, and both turned "slow" into "failed". A third change removes a race that had been letting a browser case pass for the wrong reason.

The browser's command budget. The isolated Chromium gave up on a command after fifteen seconds. On a loaded machine that is not a slow browser but a busy one. The evidence is the same runner with the same code: with a generous budget the browser panel case passed twice — the frame arrived 370ms and 1.4s after the page was opened, and the case finished in 78.7s — and with the fifteen-second budget it failed twice, both times with no frame at all. A page open on that runner takes about half a minute and a pointer dispatch takes seconds, so the old budget abandoned the open halfway: the panel never received a frame, and the person saw an empty browser instead of a slow one.

The budget is a guard against a wedged browser, so it stays a bound, but it now fits the machine. The commands that wait for the browser to work on the page get a minute; an input command — which the browser answers without waiting for the page to do anything — gets half of that, because a person is waiting for a click to land.

The verification jobs' ceiling. The WebUI group's files now run one at a time, and the Linux browser gate installs the browser's system dependencies: an apt run that took one minute in one run and thirteen minutes in the next, depending on the runner's cache. Together they put the Ubuntu job close to its twenty-five minute ceiling, and a job the runner cancels is a job that never reports. The ceiling is not a budget — a job that finishes in eleven minutes still takes eleven minutes — so every job that runs the verification pipeline now has forty minutes.

A browser case that could pass without testing what it meant to test. The agent form case asks the page for "a button named 创建 Agent" twice, without saying where that button has to be. The create menu mounts in an effect after its trigger is clicked, so the menu entry can still be missing a frame later — and the agent list carries a button of its own with the same name as soon as an agent form has been opened. Which of the two the query found therefore depended on whether the menu had mounted yet: sometimes it clicked the agent-list button and the case passed without ever going through the menu it meant to exercise, and sometimes both were on the page at once and Playwright's strict mode rejected the click. That is what turned the CLI release job red while the same case passed in the verification job two minutes earlier, on the same commit. The query is now scoped to the menu and waits for the entry to be visible, so the click lands on the menu entry every time. No expectation changed.

Validation

• pnpm --filter @mavis/webui typecheck:client — clean.
• node --test test/source-sync.test.mjs — 58 passed, 1 skipped.
• The agent form case, run eight times in a row — 8 passed. The whole plugin catalogue spec, run three times over — 15 passed.
• The platform matrix — see the checks on this pull request.

Publication and contribution checks

• I have permission to contribute these changes under the existing licenses applicable to the changed files/packages; imported material and its provenance are identified and existing notices are preserved.
• No credentials, account data, real user content, internal source history or private review material is included.
• Added/removed source files were reviewed before regenerating release/public-source.json; new tests are declared in test/vitest-suites.json where applicable.
• Shared English/Chinese documentation and capability/verification records are updated where applicable. Mock/offline results are not described as live-service acceptance.

Maintainer handoff

Publication scope or license changes (if any): none.

Shared-source port: not needed. Record public PR references only; keep private links and review material out of this PR.