[bokelley]

Fixes #1624 (part 1: display_tag inference; part 2, normalize_url whitespace, is a separate PR)

Problem

coerce_legacy_assets / infer_asset_type let an exact type-name key win over field presence. The type names came from a hand-kept list of 14 that predates AdCP 3.2, so an untyped asset under display_tag was inferred as text, url or image from its other keys:

coerce_legacy_assets({"display_tag": {"content": "<script ...>"}})
# before: {'display_tag': {..., 'asset_type': 'text'}}
# after:  {'display_tag': {..., 'asset_type': 'display_tag'}}

Fix

The vocabulary is now derived from the generated AssetVariant union (adcp.types.domains.core.assets.asset_union), the canonical creative-asset union. This follows #1501's derived format vocabulary. It is read lazily on first call and cached, so importing the legacy helpers still doesn't build the asset graph. A schema bump that adds an asset type extends key inference in the same regeneration.

Checking the full 3.2 vocabulary found 7 types missing from the hand list, not just display_tag. All are now inferred from an exact key: display_tag, zip, card, published_post, pixel_tracker, vast_tracker, daast_tracker.

Judgement call: this applies only when asset_type is absent and the key is exactly a type name. That is the precedence infer_asset_type already documents; the list had just gone stale. An untyped asset keyed exactly card or zip that used to be inferred as text/url is now inferred as its named type. That is the spec-correct reading, and the same change the issue asks for on display_tag.

Tests

• The issue's display_tag example, with both content and url assets.
• Every asset type in the format slot union's own derived vocabulary (_KNOWN_BASE_INDIVIDUAL_ASSET_TYPES) is inferred from its key, so the two generated vocabularies can't drift apart.

make lint, make typecheck-all and make test were run locally. One unrelated A2A integration test (test_a2a_context_id) flakes under local -n 8. It fails the same way on the #1619 branch and passed in #1626's CI.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — clean fix with no blocking findings.

Checked: replaces the stale 14-entry hand-kept _KNOWN_ASSET_TYPES in src/adcp/types/_legacy_assets.py with a vocabulary derived lazily and cached from the generated AssetVariant union (21 arms), so a schema bump extends key inference during regeneration. This directly addresses the repo's "hand-written vocabulary beside a generated union" concern — the derivation now tracks the generated source rather than drifting. Derivation handles both AdCPBaseModel arms and the RootModel multi-shape arms (display_tag, vast, daast, published_post); all 14 prior entries are retained with no existing-key regression. The fix(compat) semver signal is correct: the behavior change (untyped card/zip keys now infer their named type) is a documented spec-correct bug fix, not a published-contract break — no public export removed/renamed, no required↔optional flip, no enum member removal. The new cross-check test ties the legacy-derived vocabulary to format's emitted _KNOWN_BASE_INDIVIDUAL_ASSET_TYPES, preventing drift between the two generated lists.

No Critical/High/Medium findings. high_risk is false, gated_paths is false, no author team gate. Decision table rows 1–8 do not fire; falls through to row 9 (approve). Note: review_decision is REVIEW_REQUIRED, but since gated_paths is false, row 2 does not apply.