GPT-5 Codex AI Chat Assistant

Goal: make reviewed content records the source of truth for examples. Current architecture: pages still read a mix of inline examples and derived UI props. Target architecture: server code reads published records by content type and keeps compatibility only during migration. Steps: add an aggregation layer, update public pages, update audits, document compatibility behavior, then remove legacy fields after coverage. Tests: content check, media audit, model content audit, typecheck, and page sampling.

Responsibilities: compatibility definitions, admin read display, and seed support. Callers: prompt library, admin prompt page, and validation scripts. Data flow: versioned templates are the current source of truth; defaults are compatibility examples. Risky assumption: removing defaults may break labels in old scripts. Safe first change: add a usage audit before deleting compatibility exports.

Likely cause: template source exists without matching locale file. Reproduction: add slug, run build, hit prompt static params. Minimal fix: add en and zh locale files for the slug. Regression test: run prompts:check before build. Do not change: route generation unless locale files are valid.

Contract drift: schema expects image, video, audio, null, or empty compatibility. Caller impact: admin form depends on empty string handling. Validation gap: test null and empty string separately. Rollout note: normalize to null at read boundary. Risk: strict parser may reject existing draft records.