Multi-agent code review workflow
triptychworkflowsetup L3★1,223
sceneview/sceneview ↗What it does
Run pre-merge multi-agent SceneView PR review triptych
Best for
Blocking unsafe SceneView merges with enforced multi-lens coverage (correctness, API, parity, UX, safety) and a serial visual QA gate that prevents device-contention race conditions.
Inputs
- · [object Object]
Outputs
- · [object Object]
- · [object Object]
Requires
- · gh pr diff (GitHub CLI)
- · git diff (local branch)
- · bash .claude/scripts/impact-check.sh
- · Emulator/simulator (iOS/Android/Web)
Preconditions
- · 5 Opus reviewers spawn in parallel (independent, no cross-talk)
- · Updated-everywhere leg runs impact-check.sh
- · Visual device-QA runs serially (single emulator/sim lease, no parallelization)
- · All three legs complete before gating
Failure modes
- · Threading violation (Filament JNI off main thread) missed → correctness-threading lens checks explicitly
- · Cross-platform parity mismatch (Android change, iOS not mirrored) → cross-platform-parity lens flags
- · Demo has silent crash or dead control → real-product-ux lens blocks
- · New public symbol missing KDoc → api-consistency lens catches
- · .filamat ABI invariant violated → safety-security lens checks
- · Changelog missing for public change → updated-everywhere leg flags
- · Visual regression on one platform → device-QA catches (serial per platform)
Trust signals
- · 5 independent reviewers (no groupthink), each with specific lens mandate
- · Impact-check script enforces surface audit (not model memory)
- · Visual QA is human-observable (screenshots/video, not just automated tests)
- · Gate is binary CLEAR/BLOCKED (no ambiguous partial results)
- · Every blocker has file:line traceback