--- name: parked-ai-to-native-feature-audit description: "PART 1 (AI-dependence → native conversion) is CLOSED — it ran and its conversions landed. PART 2 (the full feature oversight/improvement review) may still be owed. Kept for the owner rulings it produced: FREE is the hard rule, not lightweight; and the AI→native reporting directive." metadata: node_type: project originSessionId: parked-ai-to-native-feature-audit --- **ORIGIN — owner 2026-06-29:** *"I need a full audit / review parked that checks all features that depend on AI, and makes them native where possible. Also view all features for oversights / ways to improve in that audit."* *Reduced 2026-07-20: the conversion inventory, the per-feature specs and the shipped-conversion list were removed — that work ran, and the code is the record. What follows is the reasoning and the standing rulings, which are not in the code.* ## PART 1 — AI-dependence → native conversion. **CLOSED.** The principle it was built on is permanent: **RobCo owns the truth; the AI provides perspective** ([[player-control-principle]]). Every feature that reaches for the AI gets classified — **(a) convertible to native** (deterministic, offline, no AI), **(b) must-stay-AI**, or **(c) hybrid** — and (a) gets converted. **The one MUST-STAY-AI ruling: the Director narrative.** It is inherently generative; it is not a conversion candidate. **LEAVE AS-IS:** cloud sync, auth and kill-switch flags — those are *network*-dependent, not *AI*-dependent, and already degrade offline. **The owner's own trigger for it, worth keeping verbatim:** he found the inventory USE button quietly sending a command to the AI to narrate and apply an item's effect, and asked — ***"why is use an ai command? shouldn't everything be done native?"*** He was right. **Treat that question as a standing test on any new feature.** ## ⭐ THE DURABLE OWNER RULING — FREE is the hard rule; "lightweight" is not (2026-06-30) Asked whether making image upload native was worth a multi-megabyte on-device OCR dependency: **The "lightweight / zero-byte" thing is NOT a hard rule. The owner is FINE adding a dependency if that is what it takes to make a feature native. The only HARD rule is FREE.** This overrides any inherited assumption that bundle size vetoes a native conversion. The shape that followed from it: **on-device OCR as PRIMARY with AI vision as FALLBACK** — a hybrid behind a kill-switch, not a pure replacement. And the scoping judgement that came with it: **general image *understanding* needs an on-device vision model — too heavy and too slow to be practical — but that was never the use case.** The real use is reading TEXT off game screenshots. **Don't re-open the vision-model question unless the use case actually changes.** ## ⭐ THE REPORTING DIRECTIVE (owner 2026-07-08) — still the standing pattern **When a native-conversion build runs, ALSO survey every OTHER feature that is currently AI-driven but could be native — and REPORT them to the owner. DO NOT change them. Just tell him about each candidate.** So conversion work carries a rider deliverable: a short AI→native candidate report (feature · what it currently sends the AI · whether and how it could go deterministic), surfaced for review. **Implement nothing beyond the unit's own scope until the owner picks.** He then reviews in a batch — which is how the last round went (*"approve everything you said"*). ## PART 2 — full feature oversight / improvement review. **MAY STILL BE OWED — check `QUEUE.md`.** Walk EVERY feature in the app and look for oversights, gaps, half-built bits, inconsistencies, and concrete improvement opportunities. **Produce a prioritized improvement backlog.** **Why it still has value: the owner keeps finding these live, himself** — a map marker that never persisted, a barter bug on vendor switch, overlapping stat displays, a missing native "mark visited". **Every one he finds by playing is one this audit should have found first.** ## CONSTRAINTS (apply to any resumption) - **Analysis and ideation ONLY — implement nothing until the owner reviews and picks** ([[parked-task-workflow]]: never start a parked item without an explicit go). - **Recommendations must NOT weaken the hard rules:** the full gate, notify-everything, free / BYO-key, game-agnostic. - Relates to [[parked-ai-feature-eval]] — but note they pull in **opposite directions**: that one ADDS AI features, this one REMOVES AI dependence. **Distinct goals; reconcile them in a synthesis if both ever run.**