fix: Make TestAgent and TestWorkspaceAgentPTY less flaky #1562
Conversation
| // Brief pause to reduce the likelihood that we send keystrokes while | ||
| // the shell is simultaneously sending a prompt. | ||
| time.Sleep(100 * time.Millisecond) |
There was a problem hiding this comment.
This should properly be buffered on the server-side, were you seeing it differently?
There was a problem hiding this comment.
I don't think it's a problem with the server, just an inherent race. The actual issue I was seeing was that the test was sending the command before it received the prompt, which means the command text actually got echoed twice. This is because after bash prints the prompt, it helpfully re-echoes anything that was pending in the input buffer, in order to make line editing work properly.
For instance, if you type sleep 10<ENTER>echo foo<ENTER> quickly, you get output on your terminal that looks like:
$ sleep 10
echo foo
$ echo foo
foo
It would certainly be cleaner to wait for the prompt and then send the input, but I'm not sure that we can easily/portably do that.
TestAgent/ReconnectingPTY was making some fragile assumptions about the data returned from a
net.Conn. This PR makes it a bit more robust.The previous version was sending
echo testvia a PTY, and then expecting to see the stringtestin the output twice: once from the echoed command line, and once from the command output itself. But it was expecting them to be returned by two different calls toRead(), and there's no guarantee of that.The new version uses buffered I/O to read one line at a time, and expects to see at least one line that contains
echo test, followed by at least one line that containstestbut notecho. It also sleeps for 100ms to reduce the likelihood that the shell prompt test will be interleaved with the local echo of the command.There's still some room for improvement here, because ideally the test would be controlling both ends of the PTY, instead of just assuming the user's shell behaves predictably.
Fixes #1348, fixes #1368