Claudeagent and ClaudeagentThe runner's cancellation test no longer fails on a busy machine. It sent its log to an address that is not there (127.0.0.1:9), and once a step outlasted a batch of the log (1.5 s), each batch was tried three times with waits between, about 3 s, inside the loop that times the grace after SIGINT; two batches carried the step past the 8 s the test allows, and it read as SIGINT not ending the step. The tests now use an API with no address, which sends nothing, and the log keeps what it would have sent so the test still reads it all. Failing the same way on demand (cancelling at 1.6 s instead of 0.6 s) took 10.6 s before and passes in 1.8 s now.fa14c6cHistory