McRitchie Studio

Agent Context

Follow-up to shipped #426. The live pr-review TEST (blocked-card-crew-layout) showed execution drift despite correct SOPs: Avi spawned only the primary, the primary (Shannon) then spawned the light (Carl), and the light drove the deep review + verdict (role/depth inversion). ROOT: the interactive Agent-tool path is prose-driven and drifts. FIX = make it enforced + legible: (1) Avi (supervisor) deterministically spawns BOTH primary and light in parallel as his own children; the primary must NOT spawn the light. Emit TWO intent-labeled delegate actions: 'summon primary review: <soul>' and 'summon light review: <soul>' (the Agent-tool description becomes the action label). (2) Sharpen the responsibility split in pr-review-primary.md and pr-review-light.md: PRIMARY = review owner = deep review + owns the gates (dor/cert/CI/acceptance) + DRIVES the verdict to reviewed or block. LIGHT = focused second read through its domain lens + reports to the primary; does NOT run gates, does NOT drive the verdict. (3) Add a per-domain REVIEW CHECKLIST per reviewer soul (in each soul role.md) from our hard-won gotchas: Shannon UI (Alpine x-show/single-root/x-data, Tailwind rgba space-separated vars, ERB comment rules, live-partial dual render paths, CSS-grid mb, reduced-motion), Carl backend (N+1, transactions, ErrorLog on rescue, migrations, slug FKs), Jasper web3 (IDL pin, decoder expected_len, signer order, Squads), Steffon infra, Alex docs. Domain = soul + checklist; role = primary/light responsibility. Files: bin/pr-review, docs/agents/agents/avi/sops/pr-review.md + pr-review-primary.md + pr-review-light.md, the primitive pr-review-sop.md, each soul role.md. Test the spawn structure + labels + primary-drives-verdict.