feat(ops): forward-progress watchdog — catch "busy but not progressing"

[hadamrd]

What

Adds scripts/progress-watchdog.sh — a repo-agnostic supervisor that watches forward progress (merges landing), not just liveness.

Why

Event watchers (PR_OPEN / STALL / DEAD) are blind to a loop that is busy but not progressing — spinning for hours re-reviewing one approved-but-blocked PR while every liveness check stays green. That blind spot cost ~9h of wasted grind once. The invariant that actually matters is merges landing.

Tripwires (exit non-zero so a cron/systemd/operator/agent can react)

• 10 loop entrypoint absent (DEAD) — matches the real bin/forge-loop, not the string forge-loop run (which self-matches an interactive shell)
• 11 stop-file present
• 12 two-factor stall: no merge in 90min AND the event log has been quiet 30min. A fresh event log means the loop is legitimately building/repairing a risk-gated PR that parks for human review — that is progress, not a stall, so this avoids false alarms
• 13 an open PR piling up comments past a cap (busy-but-stuck signature)

Usage

REPO_PATH=/path/to/checkout GH_REPO=owner/name scripts/progress-watchdog.sh

All thresholds are env-overridable (STALL_SECS / QUIET_SECS / POLL_SECS / COMMENT_CAP). Merge polling uses token-free git ls-remote so no GH_TOKEN is required (which would otherwise break the gh CLI).

Test

• bash -n syntax-clean
• The same logic has been running live supervising the pulsar loop, where it correctly caught a dead loop and a real risk-gated-PR review checkpoint (and the two-factor change eliminated the false 'stall' alarms the naive version produced)

🤖 Generated with Claude Code