[ideaship]

collect-kolla-logs.yml (from #2977) printed its fallback notice to stderr and looped over the stdout lines of the host discovery task. Zuul's command module merges stderr into stdout unless a job sets ansible-split-streams, and the testbed jobs don't set it. In CI the notice therefore landed in the host list.

When osism get hosts yielded nothing, the container health and log copy loops word-split the notice and ran ssh against osism, yielded, no, falling, back, to and /opt/configuration/inventory/10-custom before the real hosts. The copy loop also left a stray kolla/opt/configuration/ directory in the uploaded logs. An ssh or awk error from the manager would have been treated the same way. Seen in build cd1702dd (testbed-deploy-next-in-a-nutshell-with-tempest-ubuntu-24.04, periodic).

Change

• The discovery task writes the host list to a file on the orchestrator node. A slurp task reads it back, and the loops take their hosts from there. Stdout now carries only messages, whichever stream mode the job runs in.
• The fallback notice goes to stdout. The discovery output is printed with the copy summary, so a local run and a Zuul run report it the same way.
• The summary also runs when no host was found, so the job output shows why nothing was collected.

Testing

• yamllint and ansible-lint pass.
• I tested the file handoff alone in a stripped-down local playbook, with the notice and an ssh error line written by the discovery step:
  • With hosts, the loop gets exactly the hosts.
  • With an empty host list or a skipped discovery task, the loop is skipped.
  • In every case the notice and the error still appear in the summary.
• Not yet exercised in Zuul. The play only collects on failed builds, so a green zuul-label run would skip it; covering it needs a failing build.

🤖 Generated with Claude Code