merge-and-ci

Poll the check status via MCP on a back-off of 5s, 10s, 15s, then 30s repeating, with each sleep run in the background. Do real work or end the turn between polls instead of tight-polling or using one long sleep.

Call enable_pr_auto_merge with SQUASH so GitHub lands the PR once required checks pass. If arming is refused as unstable, keep polling and retry; if refused as already clean, squash-merge directly since the PR is ready.

The harness rejects standalone foreground sleeps like sleep 30 and requires run_in_background: true. Pass the background flag on every sleep in a polling loop starting with the first one to avoid a failed call and recovery overhead.

Usually no. A push through the in-session git proxy already triggers the test workflow on main and claude/** branches, and an open PR adds a pull_request run. Only dispatch manually when no run appeared at all.

No. Check-suite rollup events can arrive late or not at all, so a turn spent waiting on a subscription may never see the transition. Use MCP polling for merge-on-green, or arm auto-merge and end the turn.

Nothing downstream reads the working tree: capture scripts write through git plumbing against the fetched remote tip, and conformance checks resolve origin/main. A plain git fetch origin main avoids checkout denials that have cost 30-57 seconds per session.