Rui Codex

Tagline: An approval-gated agent that runs a small studio's app-store feedback loop across every app, store and language, and never posts without a human click. What it does. Rui Codex publishes 34 in-scope apps and games on Google Play and the App Store. Its feedback is a trickle scattered across two stores, nine languages and a crash console. Storefront Steward runs the developer's morning pass: a read-only store sync; triage of every review; one Muse Spark context over the whole studio, which groups reviews, crash issues and releases into verified, cross-app issues and remembers the previous pass; reply drafts in the reviewer's language, checked by a deterministic verifier; engineering tickets written from the evidence; a report. Everything waits in an approval queue. Posting is not one of the agent's tools. Its first real finding: one native crash (SoLoud::DiskFile::pos()) shared by three different games, so one fix ships to all three. Built during the hackathon (27 Oct - 4 Nov). All code in the repository was written during the building period. Built: privacy layer; budget guard and receipts; store connectors; triage, clustering and verifiers; agent and approval queue; store-screenshot pipeline; evaluation; the distilled students; keyless demo; web app. Evaluation and honesty. Two test sets, never mixed: all 49 real reviews (frozen 27 Oct), and 591 held-out synthetic reviews; 95% bootstrap intervals throughout. Gold labels are AI labels; an independent second-model review (Claude Opus 5.5, blind relabel, not a human) agreed on 50 of 50 sampled labels. Narration in the video is a disclosed synthetic voice. No real review text appears in the repository or the video. Total Meta Model API spend for the whole build: about USD 9.5. Try it without keys: python scripts/run_demo.py (synthetic reviews only).