foundation-models-auditor

Scans Foundation Models (Apple Intelligence) code for issues — both known anti-patterns like missing availability checks, concurrent requests on one session, manual JSON parsing of model output, and missing specific error catches, and architectural gaps like prompt-injection risk from direct user-text interpolation, @Generable enums without @frozen (future-case crash on iOS update), missing Cancel UX on long generations, missing transcript trimming for multi-turn chats, stale availability cache when users disable Apple Intelligence in Settings, missing retry logic for transient errors, and partial-output validation gaps.

Project Formats and Settings

The auditor reads both project.pbxproj and JSON5 project.xcproj inside .xcodeproj, including nested layouts. Select the actual project, target, configuration and SDK. It reads declarations and xcconfig files and labels unresolved values. The owning workflow can supply xcproject inspection output and a fresh matching Xcode settings capture obtained within authorized scope. An Xcode query can resolve packages or write state, so the file-read auditor does not run it. Source findings retain target membership and synchronized-folder exclusions.

What It Does

• Detects 14 known anti-patterns (no SystemLanguageModel.default.availability check before LanguageModelSession(), a second request started while the session isResponding, manual JSONDecoder of model output instead of @Generable, no catch for context-size exhaustion, no catch for guardrailViolation, session created in a button handler, no streaming for long generations, missing @Guide on @Generable properties, nested type without @Generable, no fallback UI when unavailable, no evaluation suite for a shipping AI feature, an uncalibrated model judge, an evaluation that asserts nothing, and an incomplete 27-cycle error migration)
• Identifies architectural gaps (user-controlled text interpolated directly into prompts without injection mitigation, @Generable enums not marked @frozen, missing Cancel control + Task.cancel() on streaming, no transcript trimming so multi-turn chats hit the context-window wall, Tool errors not distinguished from session errors, stale availability assumption when user toggles Apple Intelligence in Settings, streaming partial outputs displayed without empty/malformed validation, no session pooling across features, English-only error strings in localized apps, privacy-disclosure gap for on-device AI processing, optional-property output not validated, no retry on transient errors)
• Correlates findings that compound into higher severity (missing availability + no fallback UI = broken feature with no explanation, manual JSON + nested types = silent field drops, missing guardrail catch + user-controlled prompts = retry loop on safety refusal, @Generable enum without @frozen + iOS update = production-only crash on a switch fallthrough)
• Produces a Foundation Models Hardening Health Score (PRODUCTION-READY / NEEDS HARDENING / FRAGILE)

How to Use

Natural language:

• "Can you check my Foundation Models code for issues?"
• "Review my @Generable structs for correctness"
• "Audit my Apple Intelligence integration"
• "My LanguageModelSession keeps crashing"
• "Check if I'm handling Foundation Models errors properly"

Explicit command:

/axiom:audit foundation-models

Related

• foundation-models skill (axiom-ai) — on-device AI implementation patterns (iOS 26+)
• foundation-models-ref skill (axiom-ai) — complete API reference with WWDC 2025 examples
• foundation-models-diag skill (axiom-ai) — Foundation Models troubleshooting
• concurrency-auditor agent — overlaps on UI freezes from synchronous main-actor work around generation
• memory-auditor agent — overlaps on long-lived session ownership and retain cycles
• codable-auditor agent — overlaps on @Generable decode-time issues (silent field drops, future-case enum decode)
• energy-auditor agent — overlaps on repeated inference battery cost
• security-privacy-scanner agent — overlaps on user content in prompts and privacy manifest disclosure
• iap-auditor agent — overlaps when AI features are gated by purchase entitlement
• health-check agent — includes foundation-models-auditor in project-wide scans