Playwright Parallel Runs
Playwright runs share the machine lightly, and end in a result that can be trusted.
Why
Runs on a shared machine should not get in each other's way, and each should end in a result that can be trusted.
On A Shared Machine
- Every agent's Playwright run uses
--workers=1and only the touched spec. No more agents run suites at once than the machine carries without lag. Agents on default workers overload the machine and lag the director's typing. - When the machine is loaded, the agents on debt pause first.
- An agent never adds load of its own making, busy loops or CPU burners, to reproduce or steady a flaky case (Main Stays Green, Work Stays Clean).
A Full Suite
A full browser suite runs as one background command with a time limit above the suite's length, gated on its exit code. A killed run proves nothing.
Where a suite outlasts the longest a background command may run, split it in two:
- Two halves, on two ports, in two background commands at once, each under the limit and each gated on its own exit code.
- Each half gets its own output folder (
--output) and its own html report folder (PLAYWRIGHT_HTML_OUTPUT_DIR). Two runs sharing Playwright's artifact folder delete each other's traces. - The halves share one build: the first half's web server builds, and the second half gets a static server of its own on that output, started once the build is up. Two web servers building into the same folder at once collide.
- Never point a half at the other half's server, which stops when that half ends.
- A case that fails under that load and passes alone is a flaky case (Main Stays Green, Work Stays Clean).