feat(codegen): consume x-adcp-open-payload in generated models

[bokelley]

Follow-up from adcontextprotocol/adcp#5178.

The schema PR adds x-adcp-open-payload to mark intentionally open JSON payload surfaces and adds schema-authoring validation that the keyword is boolean. The Python SDK generator path should consume this intentionally rather than relying on generic additionalProperties heuristics.

Expected work:

• Add a codegen/post-processing rule for x-adcp-open-payload: true so datamodel-code-generator output preserves the intended open payload model.
• Ensure x-adcp-open-payload: false does not collapse structured-but-extension-tolerant fields into generic payload types.
• Add a smoke test using at least one annotated field from the 3.1 schema set.

Rationale: datamodel-code-generator will ignore this custom keyword by default, so without explicit SDK handling the annotation is documentation only for Python consumers.

[bokelley]

Triage

Classification: Feature request
Bucket(s): client (codegen pipeline — no codegen label exists; noted for label-gap list)
Status: deferred
Milestone: omitted — no version target in issue body

What the experts said:

• code-reviewer: Non-breaking, additive; scripts/post_generate_fixes.py is the correct insertion point using the existing schema-walk → class-injection pattern. x-adcp-open-payload: false requires only a skip/log, not a separate code path. Blocked: the keyword is absent from every schema bundle on main (including 3.1.0-beta.7), so the smoke test has no target until PR chore(protocol): update to AdCP 3.1 rc2 #901 lands and make sync-schemas re-caches the 3.1 schemas.
• ad-tech-protocol-expert: The true case is a protocol conformance requirement — without it, extension fields on open-payload surfaces are silently dropped by extra='ignore'. false should be a guard assertion in test_extra_policy.py to prevent accidental widening, not a new code path. Raises two open questions the implementation must resolve (see below).

My take: This is a clean, well-scoped follow-up — experts converge on scope and insertion point. Deferring until #901 merges; the two open questions below should be settled before the implementation PR is filed.

Blocked-on: #901 — resurfaces on merge

---

Implementation brief (for when #901 lands):

• scripts/post_generate_fixes.py: add apply_open_payload_config() — walk source schema files for x-adcp-open-payload: true, locate the corresponding generated class, inject model_config = ConfigDict(extra='allow') if absent. Follow the add_deprecated_field_metadata pattern at line ~230.
• Verify flatten_schemas() in scripts/generate_types.py does not strip the x-adcp-open-payload extension key during the JSON copy step (extension properties currently pass through; confirm this holds).
• tests/test_extra_policy.py: extend _schema_allows_extra / TestGeneratedCodeMatchesSchemas to recognize x-adcp-open-payload: true as an authoritative signal alongside additionalProperties: true; add a test_no_spurious_extra_allow guard for the false case.
• Smoke test: one fixture using an annotated field from the 3.1 schema set (requires the schemas from chore(protocol): update to AdCP 3.1 rc2 #901).

Open questions to resolve before implementation PR:

1. Does x-adcp-open-payload apply only at schema root or also on $ref-resolved sub-schemas? The _schema_allows_extra walker follows $ref chains — the new rule must match its traversal depth.
2. Precedence: when additionalProperties: true and x-adcp-open-payload: false appear on the same schema object, which wins? The post-gen rule needs an explicit tie-break.

---

Triaged by Claude Code. Session: https://claude.ai/code/cse_01QxdZunWwWsZUdsMaatDb73

---

Generated by Claude Code