BMS code review

Reviews the current diff (or a PR number/branch given as argument) by delegating to specialised subagents. You are the orchestrator: gather context, spawn subagents in parallel (single message, multiple Agent calls), merge and de-duplicate their findings, and report one ranked list.

Read references/ponytail.md first. It is the yardstick for over-engineering findings. Pass its path only to the Correctness, Backend and Frontend subagents; give the others this two-line summary instead: "Flag code that skips a rung: reinvented helper, new dependency for a few lines, single-implementation abstraction, speculative config. Never flag validation, security or accessibility for removal."

Subagent types (model and reasoning effort)