/review: add a model-family × effort routing table, starting with an A/B baseline vs a minimal prompt

Background

From a side-by-side study of /review against Claude Code's /code-review (binary 2.1.241): their command keeps an internal model family × effort level → prompt cell routing table, with each cell calibrated against an external eval set. The striking data point: on their strongest model family, medium/high effort routes to a minimal prompt (one careful senior-engineer diff pass, ≤15 findings, no angle list, no fan-out, no verify), and on the next-strongest family every level drops the verify subagents. The scaffolding's marginal value demonstrably shrinks as the model gets stronger — to the point of turning negative.

Our 14-agent pipeline was tuned on one model generation, and the routing is a constant: every model gets the same topology.

Proposal

1. First, run an A/B baseline — same PR set, --effort high full pipeline vs. a minimal prompt ("review this diff as a careful senior engineer would; ≤15 findings; every finding needs a concrete failure scenario"), compare recall and precision per model we ship on. This costs a day and produces the decision datum we currently don't have.
2. If the gap is significant, keep the pipeline and add a routing table keyed on config.getModel() family × effort, so weaker/stronger models can get different topologies (e.g. skip verification on models measured to not need it).
3. If the minimal prompt is competitive on strong models, the 60k-word SKILL.md's maintenance cost needs repricing.

Notes

• The routing table belongs in code (like parse-args' verdict), not in SKILL.md prose.
• DESIGN.md's "LLM call budget" section already tracks per-topology cost; the A/B extends it with per-model quality.

Read DESIGN.md's “LLM call budget” section, SKILL.md, and the existing /review and parse-args entry points; inspect how config.getModel() is exposed. Start with the proposed A/B comparison of the high-effort pipeline and minimal prompt across shipped models, measuring recall and precision; done means the results support either model-family routing or repricing the current pipeline.