CI on main since the flaky-test fixes

Build workflow, branch main, every run since the fixes landed on 9 Nov 2026 (12:56 UTC). No run was re-run.

6runs
3all test jobs green
3a test job failed
7tests fixed here

The runs

When (UTC)Commitubuntuwindowsmacos
9 Nov 15:34The working bar stands beside the agent's name againpasspasspass
9 Nov 16:32A web search in an agent's history shows its querypass3 failedpass
10 Nov 10:04The microphone is let go 10 s after the last recordingpasspasspass
10 Nov 10:37The phone names why Obeya does not answerpasspasspass
10 Nov 11:31The Configuration shows the usage limits in the barpass1 failed1 failed
10 Nov 14:43Obeya offers Codex's updates at a click1 failed3 failedpass

Failed without a fault in Obeya: a busy runner, or the test itself

A clone becomes the first canvas ubuntu 5.0 s, windows 25 s

Usually
0.6–0.9 s on ubuntu, 5–18 s on windows.
Cause
To read how the clone went, the test looked at the runner's whole machine. That started the Claude Code bundled with the Agent SDK, and on Windows also a PowerShell lookup. The clone itself had finished.
Now
The test sees only git and its own stand-in programs. 2.1 s → 0.2 s on this Mac.

The demo voice starts once per render windows 25 s

Usually
3.5–12 s on windows for this first test of its file, 0.3–0.9 s for the ones after it.
Cause
The first test paid for uv's first run on the runner (it has to find its Python first).
Now
That first run happens before the tests start, outside any test's time limit.

Approving directly where the work goes out as a PR macos 5.3 s

Usually
0.9–2.4 s on macos. Its neighbours took 2.4–3.0 s in that run.
Cause
One test set up two separate repositories, so it did twice its neighbours' git work within the same 5 s limit.
Now
Two tests, one per case, each with its own limit.

A removed clone's folder goes once nothing would be lost windows

Cause
Not a timeout: the test's stand-in stop command for four clones wrote to one shared file at the same moment, and Windows dropped one of the lines. Obeya had stopped all four.
Now
Each clone's stop writes a file of its own.

Failed on every Windows run (new tests that didn't work there)

Every page of obeya.si counts its visits windows, 2 of 2 runs

Cause
The test expected Unix line endings. A Windows checkout has Windows ones.
Now
It takes either.

Updating this machine's Codex (2 tests) windows, 1 of 1 run

Cause
The test wrote its search path with Unix separators, so Windows found no Codex.
Now
It uses the platform's separator.

The Windows fixes can only be confirmed by CI on main after this lands. Locally (macOS) the full suite and the type check pass.