Multi-agent teams: give the reviewer agent an explicit block-vs-file rule, or minor nits stall the critical path

In a team of ~15 coding agents with one independent reviewer agent that signs kits before runs, the reviewer slowly started blocking on minor items: extra mutant tests beyond one red per gate, log-message wording, rules for rare conditions, archive hygiene. Each block cost a full round-trip of several minutes, sometimes on the critical path.

What fixed it was a written triage rule. Block (HOLD / fix before sign) ONLY for: (a) wrong or unverifiable results, or labels that would make a claim misleading; (b) data loss or shared-host safety (OOM, locks, killing other users' processes, disk); (c) hard policy breaches; (d) a gate that can pass vacuously (e.g. a compare step that prints PASS with zero arms run). Everything else: sign, and append one line (owner, item, why minor) to a shared backlog file that owners work through when convenient.

Tips: - Ask the reviewer to re-judge its currently open blocks against the new rule and report which it downgraded. That unblocks queued work right away. - Keep the real blocks. In our case they did catch an intermittent unnamed out-of-core process and a vacuous PASS, so the rule is about triage, not weaker review.