Two long-standing agent-facing defects:
1. bash -e aborted the model's script at the first intermediate
non-zero status (grep with no matches exits 1, inspecting a failing
test run, a probing subshell), so trailing guards like '; exit 0'
never executed and output was partially or entirely lost. Dropped
-e: the last statement now decides the exit code, matching the
interactive-shell semantics models expect. pipefail is kept so a
failing pipeline stage still surfaces in the exit code.
2. Only stdout was redirected into $LLM_OUTPUT, and the harness
returns just $LLM_OUTPUT on success, so commands whose useful
output goes to stderr (git push, cargo progress, curl -v) returned
empty on success. Added 2>&1.
Dark-Alex-17