[bokelley]

Slice P1a of the reporting-subscriptions DX plan (.context/dx-reporting-subscriptions.md §6.2, P1). It adopts and serves the signed AdCP 3.3.0-beta.3 bundle. Stacked on #1751. The engine's conformance to the beta.3 rules and the re-pinned coverage vectors are in the next PR, which is stacked on this one.

Breaking change: 3.3-beta.2 is dropped

The SDK no longer advertises or accepts AdCP 3.3-beta.2. This matches how 3.3-beta.0, 3.3-beta.1 and the 3.2 release candidates were dropped once superseded.

• get_supported_adcp_versions() → ('3.0', '3.1', '3.2', '3.3-beta.3', '3.2-rc.7').
• adcp_version="3.3-beta.2" or "3.3.0-beta.2" now raises ConfigurationError.
• A peer that advertises only 3.3-beta.2 for the 3.3 line negotiates down to 3.2.
• Adopters who pinned 3.3-beta.2 (the 10.0.0-beta.3 default) should re-pin to 3.3-beta.3 or remove the pin.

The commit is feat(protocol)!: with a BREAKING CHANGE: footer. The same footer also lists the reporting-subscription removals below.

Release pin

• schemas/releases/3.3.0-beta.3.json records the four artifacts' sizes and sha256. They were cosign-verified (Verified OK) against release.yml@refs/heads/main.
• The bundle sha256 is ba61919ca0c9d67874604fbbca8945c133eaceed65da1ac962f78d94d7879f00.
• The source commit d5f8de383df13525e11721ee3d493ee76352210f comes from the Fulcio certificate (OIDs 1.3.6.1.4.1.57264.1.3 and .13). The v3.3.0-beta.3 tag (a6106149) is its Version Packages child, the same layout as beta.2.
• I synced with scripts/sync_schemas.py --require-signed-release and ran the full regeneration chain: fix_schema_refs, bundle_schemas, the versioned stubs and bases, generate_types, consolidate_exports and generate_ergonomic_coercion. I then refreshed the public API snapshot. make validate-generated is clean.

Generate Python SDK failure in the spec repo

With beta.3 schemas, the spec repo's Generate Python SDK check failed on import with PackageRequest.targeting_overlay lost its TargetingOverlayInput | None annotation. That message came from a second failure; the first one was hidden.

1. _reporting_subscription_preview raised AttributeError, because beta.3 renamed PriorVersion.
2. adcp.types.__getattr__ reported that AttributeError as a missing name.
3. The next access imported _forward_compat again, over models the first attempt had already patched.

The fix has three parts:

• The preview module is removed (see below).
• adcp.types.__getattr__ re-raises an AttributeError from the eager import as ImportError, so the first failure surfaces.
• The targeting_overlay and bundled-clone patches are idempotent.

Two subprocess regression tests cover this.

Serving the reporting-subscription fields from generated types

• adcp.types._reporting_subscription_preview is deleted. Every field it hand-modelled is generated.
• The beta.1 and beta.2 wire projections are removed: project_offering, project_reporting_delivery, project_subscription_state, project_subscription_coverage, serves_settled_subscription_fields, and the principal readback gate.
• One version predicate. adcp.reporting.subscriptions.serves_subscriptions(adcp_version) is true from SUBSCRIPTIONS_MIN_ADCP_VERSION = "3.3.0-beta.3". It replaces serves_settled_subscription_fields (beta.2), serves_inactive_coverage (beta.3) and the old service_subscriptions.serves_subscriptions (beta.1). It gates the capability block, get_reporting_status coverage (including status: "inactive" with deactivated_at) and the principal surface. ReportingSubscriptionEngine.subscription_coverage() no longer takes adcp_version.
• The schema marks the reporting-dimension vocabulary x-extensible, but the generator closes it into an enum. _forward_compat reopens the five fields that carry it (the selector, the offering, the capability block and their bundled copies) to the vocabulary's x-pattern. This is a codegen limit, not a hand-modelled field. Without it, an offering advertising standard_dimensions: ["device"] would fail to parse.

Codegen for the adcp#8547 layout in beta.3

#1741 handled the composed envelope. Beta.3 also composes core/task-submitted.json (and task-working / task-input-required) into branches and centralizes task_id and message. Three regressions showed up only on the real tree:

• Submitted arms lost status and the required task_id. CreateMediaBuySubmittedResponse(status="working") validated. The response-arm emitter now folds the composed task-state schema into each arm.
• GetTaskStatusResponse.task_id and .status became Any. Only the envelope declares them now. A pre-generation pass copies the envelope's declaration onto a root that requires it.
• About 40 task_id and message fields would have become TaskId / TaskMessage scalar classes. Callers would have had to wrap plain strings, and the adopter mypy --strict suite failed. References to those two scalar aliases now generate as str with the same constraints (min_length=1, max_length=2000). The classes are still generated.

The composed envelope on compact responses (CI follow-up)

The first CI run on this branch caught where #8547 changes behavior on the wire. These are fixed in follow-up commits:

• list_products and decline_proposals now need status. Through 3.2 their compact roots forbid it, and from 3.3.0-beta.3 the composed envelope requires it. uses_task_status_envelope(tool, version) is now version-aware, and the server, the validators and the converter pass the release. A 3.3 response is stamped status: completed, and the 3.3 → 3.2 conversion still drops it for 3.2 callers. This was the multi_platform_seller storyboard failure.
• Re-pinned tests.
  • The lattice residues are empty: every response composes both envelopes now.
  • The error-code count is 124 (CREATIVE_LIVE_UPDATE_UNSUPPORTED).
  • replayed defaults to false and accepts an explicit false.
  • The governance handlers name their decision in verdict / outcome_state, since status is the task status.
  • The round-trip corpus's check_governance response carries status.
• Webhook results carry the task status. create_mcp_webhook_payload gives a result without status the webhook's status, and replaces a model's defaulted envelope status with it, so a failed task's result never reads completed. The TypeScript SDK made the same change.
• Readback counts. core/reporting-subscription-state.json requires prior_versions_count, accounts_with_exceptions_count and accounts_unauthorized_count whenever their lists are present. The readback lists at most 16 prior versions and 1,000 accounts per list inline, each with its total. Paging the rest is in feat(reporting): apply the AdCP 3.3.0-beta.3 seller rules for subscriptions #1756.
• Checked, and fine in Python. The TypeScript SDK's generator dropped status: submitted / task_id on the proposal tools' submitted branches, and generated destination_overrides items as empty objects. Neither happens here: every submitted arm keeps Literal['submitted'] and a required task_id, and DestinationOverride has its four fields.
• Controller task_id. It narrows core/task-id.json through a one-arm allOf (maxLength: 128). The generator's merge dropped minLength: 1, so the alias now folds into the narrowing node.

Other beta.3 surface

• comply_test_controller adds reporting_subscription_probe, with operations revoke_grant, regrant, reduce_grant, set_account_timezone and advance_time. TestControllerStore gains the hook (not implemented by default) and the dispatcher validates its parameters. SCENARIOS matches the packaged schema again.
• Generated with no wiring needed: live creative updates, account-request money terms, flight_history, cancellation_effective_at, regranted_at, destination_overrides, the paged readback counts, reporting_subscription_page, offering available_from / withdrawn_at, and obligation stub.
• Generated renames: PriorVersion → ReportingSubscriptionPriorVersion, AccountsWithException → ReportingSubscriptionAccountExceptions, AccountsUnauthorizedItem → ReportingSubscriptionUnauthorizedAccount. Numbered names also renumber.
• adcp.types._generated.AllowedTask (internal) now resolves to the controller's allowed_tasks item. Both are task-name tokens. adcp.types.domains.core.AllowedTask is unchanged.

Packaging

• The wheel and sdist ship _schemas/3.3.0-beta.3 and _compliance/3.3.0-beta.3.
• The beta.2 cache and compliance trees stay in the repo as history and are pruned from distributions.

Tests and docs

• The adoption test is renamed to test_adcp_3_3_0_beta_3_adoption.py. It checks the signed identity values: 4684 manifest files, 247 provenance files and 1768 schemas. It also checks that beta.0, beta.1 and beta.2 are rejected.
• The version sweep, the 3.3 downgrade test, and the packaging and historical tests point at the new pin.
• The reporting tests that asserted beta.1 and beta.2 projections now assert the unprojected beta.3 shape.
• MIGRATION_v9_to_v10.md has the pin, the breaking note and rows for the new surface. docs/adcp-3.3-sdk-inventory.md has a new §13. Pin references are updated in RELEASING.md, docs/cross-version-conversion.md, docs/agent-identity-3-3-migration.md and docs/reliable-reporting-service.md.

Proof

>>> adcp.get_adcp_version()                                   # deprecated alias of the spec pin
'3.3.0-beta.3'
>>> adcp._version.resolve_adcp_version(None)                  # wire pin
'3.3-beta.3'
>>> adcp._version.get_supported_adcp_versions()
('3.0', '3.1', '3.2', '3.3-beta.3', '3.2-rc.7')
>>> negotiate_version(["3.2", "3.3-beta.3"], maximum="3.3-beta.3")
'3.3-beta.3'
>>> negotiate_version(["3.2", "3.3-beta.2"], maximum="3.3-beta.3")   # superseded beta
'3.2'
>>> ADCPClient(AgentConfig(id="a", agent_uri="https://example.com/mcp", protocol="mcp")).adcp_version
'3.3-beta.3'

Local verification

• These pass: ruff check src/, mypy src/adcp/, the mypy --strict adopter checks, the pyright adopter checks, the type-ignore contract, the versioned stub and base checks, and make validate-generated.
• Focused suites pass: 1,006 tests covering codegen, the version pin, the public API, the adoption test, server DX, targeting_overlay compatibility, every reporting-subscription and principal test, and lazy types. Codegen, version and historical tests add another 1,236.
• The full suites run in CI (dispatched manually, since stacked PRs don't trigger ci.yml).

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — PR #1754 adopts the signed AdCP 3.3.0-beta.3 bundle with a correctly-signaled breaking change.

The reviewer surfaced no critical, high, or medium findings. Walking the decision table top-down:

• Row 1 (critical/high): no such findings — not triggered.
• Row 2 (gated_paths): gated_paths is false — not triggered.
• Row 3 (high-risk deletion): high_risk is true but every high-risk reason is (added) — new schemas/cache/3.3.0-beta.3/** files (scaffolding of the new schema bundle), no (deleted) entries — not triggered.
• Row 4 (medium + data-loss/schema/infra): no medium findings — not triggered.
• Row 5 (high-risk modified + medium): all high-risk matches are additions and there are no medium findings — not triggered.
• Row 6 (sticky escalate): no prior decision — not triggered.
• Row 7 (no-auto-approve team): no team match — not triggered.
• Row 8 (≥3 medium): zero medium findings — not triggered.
• Row 9: approve.

The high_risk flag fires only on newly-added schema cache files, which is normal bundle-scaffolding and not escalation-worthy per the change-kind rules. The reviewer verified the load-bearing hand-written changes: version-aware uses_task_status_envelope, idempotent _forward_compat.py patchers, the eager-import ImportError re-raise, the breaking serves_subscriptions collapse, webhook task-status stamping, and required *_count fields — all covered by tests. The breaking change carries feat(protocol)!: + a BREAKING CHANGE footer plus MIGRATION_v9_to_v10.md and a refreshed public_api_snapshot.json, satisfying the semver-signal mandate.

No blocking findings. Ready to merge.