Principle Non Blocking Waits

Non-Blocking Waits

Background external waits—CI, reviewers, vendor CLIs, long jobs—in one call ending with the awaited task; never foreground sleep or poll tasks whose completion the harness reports.

• Use one Bash call with run_in_background: true; its completion notification wakes the session without a foreground ceiling.
• Put each poll in that call: sleep <interval>; <poll command>; use one call per bounded cycle.
• Never poll a backgrounded task with loops such as sleep 120; wc -c <output-file>.
• Apply principle-bounded-loops: declare the cap and make exhaustion loud and terminal.
• If background execution is unavailable, state the fallback at the call site and chunk foreground waits below the harness ceiling.
• Wait inline only below roughly a minute, such as a few seconds after push; background waits that would otherwise fragment.
• Preserve the constraint: Claude Code kills foreground Bash at 600 s; timeout 1800 still dies after ten minutes with exit 143.
• Never size around that ceiling with sleep 570, sleep 590, or sleep 600; a 24-hour foreground watch costs ~192 turns and re-sends context.