[bokelley]

Closes #1560. Follows the main → next sync (#1578) and the inventory (#1660). Squash-merge, keeping the ! and the BREAKING CHANGE: footer from the commit message (below).

The bundle

• schemas/releases/3.3.0-beta.0.json pins the signed release: source commit a2d7eac350f0325331ae544bc071301afe01d2f2, identity release.yml@refs/heads/main, four artifacts with sizes and sha256. Verified locally as the release notes prescribe: sha256sum -c OK and cosign verify-blob "Verified OK" against that identity; scripts/sync_schemas.py --require-signed-release re-verifies the same four artifacts and the signature on every sync.
• src/adcp/ADCP_VERSION is 3.3.0-beta.0; the bundle key keeps the full identifier (schemas/cache/3.3.0-beta.0, src/adcp/_schemas/3.3.0-beta.0, src/adcp/_compliance/3.3.0-beta.0: 4443 manifest files, 242 provenance files), with MANIFEST.in and package-data entries, so the protocol repo's "Generate Python SDK" check finds the pinned key. The signed 3.2.4 bytes stay installed under schemas/cache/3.2 as a preview bundle (PREVIEW_VERSIONS = ("3.2.4", "3.1.27")) for negotiation and translation with 3.2 peers.
• Regeneration: 181 generated files change and 44 are added (1135 total). The delta is semantic (SCHEMA_DELTAS.md); the numbered-class churn in it is a consequence of new sibling schemas and cannot be discarded. No hand-written module referenced a removed numbered name. make validate-generated is deterministic on the new tree.
• From the bundle: skills/adcp-media-buy/SKILL.md, the request-signing conformance vectors (130 files, manifest pinned), the 3.3 compliance package.

Version negotiation and translation

• COMPATIBLE_ADCP_VERSIONS adds 3.3; get_supported_adcp_versions() advertises the exact pre-release 3.3-beta.0 (never a stable 3.3): 3.0, 3.1, 3.2, 3.3-beta.0, 3.2-rc.7.
• adcp.compat.versions converts the 3.3 line: the SDK's served pre-release is its line's representative (_wire_for_line), so 3.3-beta.0 ↔ 3.2 ↔ 3.1 ↔ 3.0 step through the bundled schemas; any other pre-release keeps its own contract (is_convertible_version). Downgrades still go through _require_target_valid (3.2 to 3.1 conversion emits payloads 3.1 refuses: empty notification event_types, and geo_regions as {} #1621 / fix(compat): never downgrade into a payload the older schema refuses #1630): a converted payload the older schema refuses fails as ADCPVersionConversionError. The version envelope is stamped only where the target contract admits it: 3.3 added adcp_version to list_products responses, whose 3.2 schema forbids it.
• The client negotiates from the served pre-release (_convert_outgoing_request no longer treats it as a historical pin), and a server_version pin keeps wire precision ("3.1.0-beta.1" → "3.1-beta.1"), so a pre-release pin is never echoed as its bundle key.
• A server that restricts supported_versions without adcp_version pins to its newest listed release (version_policy.default_adcp_version) instead of refusing because the SDK default moved past it. The decisioning capabilities gate (startswith("3.2")) is now "3.2 or later".
• Signing: the 3.3 line maps to the 3.2 RFC 9421 profile; rfc9421_deprecation_active() is now true by default, so legacy SigningConfig / RequestSignatureVerification construction warns, as 3.3 deprecates those profiles.

Forward-compat patches retired (each verified against the 3.3 schema)

• reporting_webhook.operation_id (adcp#7895): core/reporting-webhook.json declares it with minLength 1, maxLength 255 and the same pattern, optional, never null. The _forward_compat.py runtime patch and _ReportingOperationId are deleted; ReportingWebhook and the request facades bind the generated classes. tests/test_reporting_webhook_operation_id.py passes unchanged.
• list_creatives creatives[].platform_id (adcp#8175): native in 3.3 (Creative.platform_id). This SDK carried no forward-compat patch for it, so there was nothing to retire.
• Still needed: the canonical_creatives capability patch (3.3 does not declare it), the targeting-overlay rc.3 bridge and the readback manifest patches.

Identity preview retired

The vendored identity schemas (src/adcp/signing/_identity_schemas/, adcp main 277e0a5) are semantically identical to the bundle's (only $ref prefixes differ), so adcp.types._identity_preview, the vendored tree and scripts/generate_identity_preview.py are deleted. adcp.types.experimental_identity keeps every exported name on the generated 3.3 types; AgentOnboardingRequirement.kind stays open (enum | str, the schema enum is x-extensible) through inherit_field redeclarations down the capabilities tree, with one exemption recorded in tests/test_redeclared_fields_inherit.py. adcp.signing.trust validates trust records and requirement kinds against the bundle through get_named_validator / get_named_schema_document; identity_schema_bundle() is gone.

3.3 task surface

list_collections and list_proposals are wired through ADCPClient, ADCPMultiAgentClient, ADCPHandler (not-supported defaults), the MCP and A2A adapters, the CLI dispatch and the MCP tool definitions, the way list_products is. The test controller gains the 3.3 simulate_delivery params (property_delivery, installment_delivery, installment_property_delivery) and the four new scenarios, raising NotImplementedError like every other unimplemented scenario. sales-fixed-rate joins the specialism enum and maps.

Generator

• Scalar roots whose pattern needs lookaround are generated with model_config = ConfigDict(regex_engine="python-re"); _scalar_root_spec now keeps that engine as a constraint, so PositiveDuration and Domain become ScalarStr subclasses like every other scalar root.
• UniversalMacro is a RootModel union in 3.3 (enum | GDPR_CONSENT_<gvl_id> pattern); the canonical direct scan reads scalar-root and root-model values as the dump would instead of falling back to a full dump.

Reporting

The production and feed code derived wire versions correctly once the client stopped echoing bundle keys; feed/snapshot.py's unversioned-walk branch accepts any served release at or above 3.2-rc.6 instead of a literal 3.2 set. The production reporting mount still advertises exactly its pin, so an SDK 10 mount on the default pin refuses 3.2 buyers unless constructed with adcp_version="3.2" (flagged for workstream 3 in the inventory). The receipt-submission plan stays tagged 3.2 (frozen bytes).

Breaking changes (see MIGRATION_v9_to_v10.md)

Wire pin 3.3.0-beta.0 and default adcp_version: "3.3-beta.0"; UniversalMacro union (UniversalMacro3 is the enum); the identity preview modules removed and experimental_identity bound to generated types (RequestSigning.profiles: list[Profile], PrincipalChangedWebhook.reason is the generated enum); identity_schema_bundle() removed; RFC 9421 construction warns; evaluator_id gone from evaluator specs; server_version pins keep wire precision; schema defaults appear in dumps for the new 3.3 fields (include_collections, hold_extension); delivery rows may omit impressions.

For workstream 2 (first items)

• Schema defaults materialized on dump (model_dump emits schema defaults the wire document omitted (143 fields in 60 schemas), which changes digested and echoed bytes #1613) now have user-visible cost: every get_products request carries include_collections: false, which a 3.3 → 3.2 downgrade drops with a recorded loss and a buyer-side warning per request, and criteria.product_ids_exclusive: false had to be special-cased on the legacy catalog route. Dumping outgoing requests with "what the caller set" semantics would remove the class.
• x-extensible enums generate closed (agent-onboarding-requirement-kind, advertiser-industry, media-buy-action-mode); the identity facade widens one by hand.
• No v33 versioned view modules yet (scripts/generate_versioned_stubs.VERSIONS).

Validation (local, Python 3.12)

make lint, make typecheck-all, make validate-generated: pass. Full suite (pytest tests/ -n auto): 19073 passed, 2730 skipped; the 12 failures and 1 error of the parallel run (a2a context id, freezegun fresh process, versioned-bases mypy subprocess, two staging-reuse writer tests, a notification-outbox import test, plus the four fixed here) all pass serially, as they did on #1578. Postgres lanes are skipped locally.

BREAKING CHANGE: the wire pin is AdCP 3.3.0-beta.0 and the default adcp_version is "3.3-beta.0" (pin adcp_version="3.2" for a 3.2 contract); adcp.types.UniversalMacro is a RootModel union whose closed enum is UniversalMacro3; the identity preview modules (adcp.types._identity_preview, adcp.signing._identity_schemas, scripts/generate_identity_preview.py) are removed and adcp.types.experimental_identity binds the generated 3.3 types (RequestSigning.profiles is list[Profile], PrincipalChangedWebhook.reason is the generated enum); adcp.signing.identity_schema_bundle is removed; legacy RFC 9421 SigningConfig and RequestSignatureVerification construction emits a DeprecationWarning; evaluator_id is gone from evaluator specs; server_version pins keep wire precision; schema defaults for the new 3.3 fields appear in dumps. See MIGRATION_v9_to_v10.md.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon is not auto-reviewing this PR because it modifies the review system itself (LADON.md). A human reviewer should review and merge this PR; Ladon resumes on subsequent PRs once these changes land on main.

[aao-secretariat]

Ladon verdict: Approve

Approve — packaging/docs/schema-bundle pass with no blocking findings.

This pass is confined to docs, packaging (MANIFEST.in, pyproject.toml), auto-generated shared-type-names.md, and the pinned signed AdCP 3.3.0-beta.0 bundle bytes. No hand-written source changed; the source wiring was prior-reviewed. Packaging is complete and consistent: the new 3.3.0-beta.0 cache/_schemas/_compliance trees are added to both MANIFEST.in and package-data, and retired _identity_schemas/_identity_preview entries are removed from both package-data and the ruff exclude.

Decision-table walk: the high_risk flag is true only because many new schemas/** files are added (plus three renames) — all (added)/(renamed), no (deleted) and no modified high-risk file carrying a medium finding, so rows 3 and 5 do not fire. gated_paths is false and no no-auto-approve team matches. No critical/high/medium findings. Only two Low doc-accuracy nits recorded, which are omitted from blocking but noted below. Falls through to row 9 → approve.

Nits (non-blocking):

• docs/adcp-3.3-sdk-inventory.md:143 — inventory status for list_collections/list_proposals contradicts PR body's "wired" claim.
• MIGRATION_v9_to_v10.md:53 — PositiveDuration/Domain RootModel→ScalarStr (.root removed) shape change not mentioned in the migration note (semver signal otherwise present via feat!/BREAKING CHANGE).

[aao-secretariat]

Ladon verdict: Approve

Approve — AdCP 3.3.0-beta.0 bundle adoption, subsequent pass with no actionable findings.

The delta is overwhelmingly generated artifacts (schema cache under schemas/cache/3.3.0-beta.0/**, domain types, conformance vectors). The reviewer verified the hand-written layer:

• New list_collections/list_proposals wiring across client/base/mcp/a2a/builder/mcp_tools/types — consistent and complete.
• compat/versions.py translation (_admits_version_envelope safe; is_convertible_version matches documented prerelease design).
• _forward_compat.py operation_id retirement verified against 3.3 schema constraints.
• decisioning/accounts.py loader fail-closes on id mismatch with multi-tenant scoping documented.
• post_generate_fixes.py regex_engine preservation intact.
• Semver signal present (feat(protocol)! + BREAKING CHANGE footer + MIGRATION_v9_to_v10.md); pathway-back rule satisfied (product_ids_exclusive refused, dropped default recorded as loss).

Decision-table walk: no critical/high/medium findings (row 1 no). gated_paths=false (row 2 no) and review_decision is APPROVED regardless. High-risk flag is true but all reasons are (added) scaffolding or (renamed) codegen churn — no (deleted) (row 3 no), no medium findings to pair with (rows 4/5 no). Prior decision was approve, not escalate (row 6 no). No no-auto-approve team match (row 7 no). Zero medium findings (row 8 no). Falls through to row 9 — approve. Clean net-improvement iteration on top of a prior clean approve.

[aao-secretariat]

Ladon cannot review this PR until merge conflicts are resolved.

[aao-secretariat]

Ladon verdict: Approve

Approve — AdCP 3.3.0-beta.0 migration delta, clean pass.

The reviewer found no critical/high/medium findings. Checked:

• Walked the decision table top-down. No critical/high findings (row 1 no), gated_paths is false (row 2 no), no (deleted) entries in high_risk_reasons — only (added) and (renamed) (rows 3 no). No medium findings at all, so rows 4/5/8 don't fire. Prior decision was approve, not escalate, so row 6 doesn't apply. No no-auto-approve team match (row 7 no). Falls through to row 9 → approve.
• high_risk: true is driven entirely by added/renamed schemas/cache/** codegen artifacts — new-file scaffolding and traversal-order renumbering, which the rules explicitly treat as normal non-escalation-worthy churn when the reviewer surfaces no medium-or-higher concerns.
• Hand-written logic (version-line translation, brand_jwks bounds/timeouts, JSON bounds scanner, webhook_receiver, forward-compat retirement, pixel_tracker) all verified clean; breaking public-surface changes carry the feat!/BREAKING CHANGE semver signal; video_brief routes correctly in both asset unions and vocabularies.

No blocking findings, no medium findings, no escalation trigger.

[aao-secretariat]

Ladon verdict: Approve

Approve — clean pass on PR #1662.

No critical/high/medium findings. Walking the decision table top-down:

• Row 1 (critical/high): no findings — skip.
• Row 2 (gated_paths unsatisfied): gated_paths is false, and review_decision is APPROVED anyway — skip.
• Row 3 (high-risk deleted): high_risk is true, but all high_risk_reasons entries are (added) or (renamed); no (deleted) — skip. The renamed core/agent-onboarding* entries are treated as modifications, not deletions, and carried no findings.
• Row 4 (medium data-loss/schema/infra): no findings — skip.
• Row 5 (high-risk modified + medium): no medium findings — skip.
• Row 6 (sticky escalate): prior decision was approve, not escalate — skip.
• Row 7 (no-auto-approve team): no team match — skip.
• Row 8 (≥3 medium): zero medium — skip.
• Row 9: approve.

The high_risk flag fires only because the diff adds the 3.3.0-beta.0 schema cache tree (schemas/**), which is normal scaffolding/regeneration output — not an escalation trigger. The reviewer verified the hand-written delta: the ADCP_VERSION 3.2.4 → 3.3.0-beta.0 bump carries the correct feat(protocol)!: breaking marker plus MIGRATION_v9_to_v10.md; new list_collections/list_proposals tasks are fully wired and snapshot-tested; cross-version translation helpers are covered at both transport boundaries down to 3.0/3.1; _forward_compat.py changes and the open asset-union vocabulary check out. Prior pass was clean and remains so.

[aao-secretariat]

Ladon verdict: Approve

Approve — clean 3.3.0-beta.0 adoption with no findings.

No critical/high/medium findings surfaced. The delta since the prior approve is a rebase + full regeneration (gen date 2026-10-26); the ClassNameN renumber churn is codegen, not a schema delta.

Checks verified by the reviewer:

• Load-bearing version-translation logic (compat/versions.py _wire_for_line/is_convertible_version/_admits_version_envelope) correctly steps 3.3-beta.0↔3.2↔3.1↔3.0 and drops adcp_version only where the target contract forbids it (list_products), with matching cross-version + logging tests — DR-0032/DR-0033 pathway coverage present.
• signing/autosign.py clamps 3.3 to the 3.2 RFC 9421 profile.
• reporting_webhook.operation_id / identity-preview forward-compat retirements are backed by the generated 3.3 class carrying the same constraints + NoExplicitNull — no fallback arm lost.
• Semver signal present (feat!: + BREAKING CHANGE squash, MIGRATION_v9_to_v10.md).
• No generated-tree hand-edits, no credentials in ctx_metadata, import layering intact, CI gates armed.

Gating posture: high_risk is true, but all schema matches are (added) or (renamed) (treated as modification) scaffolding/regeneration with no medium-or-higher finding on them, so rows 3–5 do not fire. gated_paths is false and review_decision is APPROVED. No no-auto-approve team match. Prior decision was approve, so sticky escalation (row 6) does not apply. No findings → rows 1, 8 do not fire. Falls through to row 9 = approve.

[aao-secretariat]

Ladon verdict: Approve

Approve — subsequent clean pass on the AdCP 3.3 upgrade.

No critical/high/medium findings. Walking the decision table top-down:

• Row 1 (critical/high): no such findings.
• Row 2 (gated_paths): gated_paths is false; also review_decision is APPROVED. Not triggered.
• Row 3 (high_risk + deleted): high_risk is true, but the high_risk_reasons entries are all (added) or (renamed) — no (deleted). Renames of a few core schema files are codegen churn, not destructive. Not triggered.
• Row 4 (medium + data-loss/schema/infra): no medium findings.
• Row 5 (high_risk modified + medium): no (modified) entries carry a medium finding (no mediums at all). Not triggered.
• Row 6 (sticky escalate): prior decision was approve, not escalate. Not triggered.
• Row 7 (no-auto-approve team): no team match.
• Row 8 (≥3 mediums): zero mediums.
• Row 9: approve.

The high-risk flag fires only because the diff touches schemas/** cache files — these are added/renamed generated schema cache artifacts (AdCP 3.3.0-beta.0), normal scaffolding/regeneration output, not escalation-worthy on the flag alone.

Reviewer verified the load-bearing contracts: semver signal (9→10 major carries ! + BREAKING CHANGE + MIGRATION note), forward-compat unions (no fallback arm removed, discriminator intact), the version-translation pathway (droppable/refused/moves handled with tests), and A2A fail-closed hardening. One Low nit in scripts/post_generate_fixes.py (unused extra_constraints) is non-blocking and omitted from the gate.