Steven Rostedt told the list that ktest.pl, in tools/testing/ktest/ in the kernel tree, parses git's output to show the first bad commit in an automated bisect. The recent change 0c0f93e7fa64 (bisect: print bisect terms in single quotes) made it fail to find the commit. A kernel-side fix is already posted, and he wrote mainly so the Git developers know about the dependency.

Rostedt also showed that the script matches the "Bisecting: ... (roughly N steps)" line and the "<sha> is the first bad commit" line from the last bisect action, to decide whether the bisect has finished. He asked whether Git has a command to tell that a bad commit has been found. In reply to Kristoffer Haugsbakk, he said ktest.pl builds, installs, boots and tests each kernel, so it does much more than "git bisect run".

Theodore Tso said "git bisect run" is not always the right tool. His gce-xfstests setup drives bisection across several cloud VMs. He checked his own test appliance and found it depends only on whether the output contains "is the first bad commit". He suggested a "git bisect --format=json" option for programs that run bisect in automation, and noted that a commit summary containing that phrase, a change to the text, or a non-English LANG setting would break such parsing.

Junio C Hamano asked whether the script should instead look at "git show bisect/bad". Kristoffer Haugsbakk pointed to recent threads on querying for the first bad commit, and to an unfinished script of his own that used exit codes, ref state and possibly git rev-list --bisect.

Rostedt said he could look into changing ktest.pl, which dates to 2010. No Git-side change has been proposed yet.