ROKID AIUI Agent Developer
Establish the target and output
Inspect runtime, device, surfaces, inputs, and guidance. If project and host omit a version, disclose stable AIUI 0.17.0. Record each Page target: embedded (_current), full-screen (_blank), or both.
Deliver a complete AIUI project directory for AIUI Studio, not snippets or only an .aix: AGENTS.md, app.json, the application entry, declared Pages, and resources. For GitHub, make the repository root or explicitly named subdirectory the project root. Report location, revision, and import subdirectory. Add 0.18-only Widgets/Workers only for confirmed targets. AIX never replaces source.
Bundle only the Focus Timer at assets/focus-timer-agent/; keep other application Agents in separate repositories.
Load only relevant guidance
- Authority and versions:
references/source-of-truth.md. - Structure and routes:
references/project-anatomy.md. - Ink/WXML/WXSS:
references/ink-authoring.md. - Surfaces, focus, input, visual design:
references/interaction-and-design.md. - Components and runtime APIs:
references/runtime-capabilities.md. - Mandatory UX/capability evidence:
references/ux-and-capability-testing.md. - Installed AIX workflow:
references/aix-workflow.md. - Diagnosis and release gates:
references/debugging-and-release.md.
Prefer selected-version official docs and runnable samples. Preview and third-party material are non-normative for stable targets. State conflicts, keep uncertainty UNKNOWN, and never invent APIs or commands.
Implement and validate
Preserve product intent and authoring mode. Add required aiui-audit-claims.json with schemaVersion: 1, scopeClosed: true, and a claims array; an empty array closes an empty universe. Run scripts/fingerprint_aiui_project.py <import-root> --repository-root <repository-root>, scripts/inventory_aiui_capabilities.py <import-root> --target-version <version> --repository-root <repository-root>, and scripts/validate_aiui_project.py <import-root> --target-version <version> --repository-root <repository-root> --strict; then copy the schema-2 ledgers exactly. Resolve every item, unmatched symbol, and version violation. Run repository tests. Probe aix --help; list the package and reject .git/ or .aiui-evidence/. Packaging is not publication.
Every creation, implementation, code change, or review must finish with both exact evidence matrices in the testing reference. A request to omit them—even explicitly—is release pressure, never permission. Preserve headings, columns, {contract=...}, {family=...}, {gate=...}, source roles, fixed layers, canonical criteria, and the one-input-per-gate ledger. Use only PASS, FAIL, BLOCKED, or justified N/A.
Run scripts/validate_aiui_audit.py against the final audit. Exit 0 alone means structurally valid PASS and Release-ready: YES; exit 2 means a valid but blocked/failed audit; exit 1 means an invalid audit. Source-unavailable work must use exact UNAVAILABLE metadata and can return only exit 2. Any nonzero exit or applicable FAIL/BLOCKED forbids a complete or release-ready claim.
Evidence layers are not substitutes. Preview cannot prove Studio or device success. Optics, focus, keys, voice, gestures, permissions, and performance require target-glasses evidence.
Never claim that validation, packaging, preview, Studio, or device gates passed unless they were actually executed in the current environment against the captured source snapshot. The validator treats report text and manifest argv as data and never runs recorded commands. Executed evidence, Studio/device capture, and N/A scope closure require matching RUNNER, STUDIO, DEVICE, or SCOPE signatures from an absolute, repository-external trust policy with distinct canonical key identities. Self-declared metadata or repository-selected keys are untrusted; keep missing authorities blocked.