Regression in 6.19.0: per-asset-type requirements codegen as z.any() (was typed in 6.18)

[bokelley]

Summary

Following the great work in #1654 (named slot unions + Format.assets typed), 6.19.0 introduces a runtime-Zod regression on the per-asset-type requirements bag. In 6.18.0 the codegen produced typed Zod schemas (ImageAssetRequirementsSchema, TextAssetRequirementsSchema, …); in 6.19.0 those Zod schemas are gone, and the requirements field inside each Individual<Type>AssetSchema is now z.optional(z.any()).

The TS types for the requirement bags are unchanged (ImageAssetRequirements, TextAssetRequirements, etc. still live in core.generated.ts and are referenced correctly by Individual<Type>Asset interfaces). The regression is only on the runtime Zod side.

Evidence

From @adcp/sdk@6.19.0/dist/lib/types/schemas.generated.d.ts:

export declare const IndividualImageAssetSchema: z.ZodIntersection<
  z.ZodObject<{ /* base slot fields */ }, z.core.$loose>,
  z.ZodObject<{
    asset_type: z.ZodLiteral<"image">;
    requirements: z.ZodOptional<z.ZodAny>;   // ← was ImageAssetRequirementsSchema in 6.18
  }, z.core.$loose>
>;

export declare const AssetRequirementsSchema: z.ZodUnion<readonly [
  z.ZodAny, z.ZodAny, z.ZodAny, z.ZodAny, z.ZodAny, z.ZodAny,
  z.ZodAny, z.ZodAny, z.ZodAny, z.ZodAny, z.ZodAny, z.ZodAny
]>;

Compare 6.18, where TextAssetRequirementsSchema (and the 11 siblings) were fully typed Zod objects with min_length, max_length, character_pattern, prohibited_terms, etc., and AssetRequirementsSchema was a proper z.ZodUnion of those 12 objects.

The per-asset-type *AssetRequirementsSchema exports are also gone from 6.19 (replaced by the Individual<Type>AssetSchema slot wrappers). Downstream consumers that previously imported them — including agentic-api in scope3data/agentic-api#2366 — fail to resolve.

Impact

• Slot-level schemas no longer validate the shape of requirements at runtime. {asset_type: "text", requirements: { max_animation_duration_ms: 5000 }} parses cleanly even though max_animation_duration_ms is an image-only field.
• AssetRequirementsSchema is unusable as a standalone validator — it's structurally z.any().
• Consumers that adopted feat(types): tighten Format.assets typing + emit named slot unions/Zod schemas #1654's slot unions lose the field-level validation they got from feat(types): tighten Format.assets typing + emit named slot unions/Zod schemas #1654's predecessor.

Likely cause

Speculation, but the slot-level intersection pattern in #1654 may have tripped the ts-to-zod codegen on the per-type requirement objects, falling back to z.any() for the requirements field. The TS types were unaffected because they don't go through the same codegen path.

Ask

Restore typed Zod for requirements in each Individual<Type>AssetSchema (and the parallel Group<Type>AssetSchema). Either by re-exporting the per-type *AssetRequirementsSchema and referencing them inside the slot schemas, or by inlining the requirement shape directly in each slot schema's intersection.

Cross-link: AdCP protocol context adcontextprotocol/adcp#3305 (Creative Formats v2 RFC, validate_input scaffolding).

Versions

@adcp/sdk@6.19.0 exhibits the regression. @adcp/sdk@6.18.0 does not — those typed schemas existed there.

[bokelley]

Triage

Classification: Bug (regression)
Bucket(s): library, schema-sync (label gap: schema-sync not in repo)
Status: drafting-pr (PR #1660 already addresses this)

What the experts said:

• code-reviewer: Root cause confirmed — import type {…} from './core.generated' injected by applyIndividualAssetDiscriminators appears mid-string in the combined sourceText passed to ts-to-zod; ts-to-zod can't resolve the path, emits non-exported z.any() stubs that overwrite the correct schemas from coreContent. Fix: strip the import from toolsContent before concatenation using a gm regex with a strip-count guard.
• ad-tech-protocol-expert: *AssetRequirements shapes are spec-derived (pulled from schemas/cache/latest/ via sync-schemas); the fix belongs entirely in generate-zod-from-ts.ts, no spec-repo action needed. Restoring typed Zod correctly reflects the spec's intent for per-asset-type requirements.

My take: PR #1660 ships the right fix (same root cause, same strip approach). The concurrency window meant triage started before the PR existed. Both IndividualAssetSchema and GroupAssetSchema slots are covered once the import strip lands; AssetRequirementsSchema is restored to a real discriminated union.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_013K5WTtFH1p6KddJJ2D5KHK

---

Generated by Claude Code