[Synaeve]

Refs adcontextprotocol/adcp#8534

Problem

buildComplianceBundleResults grades a storyboard whose only result is the runner's root capability_unsupported row as not_applicable. This is the rule from #2960: an optional scenario the agent never claimed does not cap its bundle. The rule applies only when the storyboard's loader-stamped adcp_version is 3.2 or later (CAPABILITY_ROLLUP_MIN_ADCP_VERSION = '3.2'). #2960 deferred 3.1 for two reasons. The 3.1 governance scenarios lacked a capability predicate, and every 3.1 release lacked the "Applicability order" rule that makes this grade normative.

adcontextprotocol/adcp#7625 (closing adcp#7584) fixed both on the 3.1 line. It backported the rule, and it gated the governance scenarios on media_buy.governance_aware. adcp#7584 asked for this explicitly "so adcp-client#2960's rule can apply to 3.1 bundles as well". Every stable 3.1 bundle from 3.1.24 carries the same text as 3.2: an unsatisfied root capability predicate "keeps optional capability storyboards out of the coverage totals of agents that never claimed the capability". Other 3.1 scenarios already declared root predicates before 3.1.24; what they lacked was the rule. #3520 already lets these gated storyboards reach the runner, so only the rollup cutoff was left.

Today a 3.1 seller that declines optional capabilities keeps partial bundles. For the non-guaranteed media-buy seller in adcp#8534, graded on 3.1.29, 31 declined scenarios keep the media-buy protocol bundle partial. The universal wholesale_feed_* bundles stay partial too. Strict Spec grading on agenticadvertising.org reads these bundle verdicts.

Changes

src/lib/testing/compliance/comply.ts

• isCapabilityUnsupportedSkip now asks carriesApplicabilityOrderRule(adcp_version). That is true from 3.2, as before, or for a stable 3.1.N release with N >= 24.
• compareAdcpVersionStrings orders by major.minor and ignores the patch, so the 3.1 line is matched on its exact patch (STABLE_3_1_RELEASE_PATTERN). 3.1, v3.1.24 and 3.1 prereleases still fail closed.
• Every other condition is unchanged: the declared predicate, the compliance_testing.* refusal, the result/phase/step allowlists, the counts and the empty skip.requirement.
• 3.0 and 3.1.0–3.1.23 bundles keep capping.

src/lib/testing/storyboard/runner.ts

Comment only: the applicability order is noted as backported to 3.1.24.

test/lib/storyboard-capability-rollup.test.js

• The version table now pins 3.1.23, 3.1, a 3.1 prerelease and v3.1.24 as partial, and 3.1.24 and 3.1.31 as passing.
• The real-cache test now checks every capability-gated scenario of the 3.1.31 bundle (not compliance_testing.*). Each one is neutralized as loaded, and stays partial when re-stamped 3.1.23.
• The header and legacy-fixture comments now describe the backport. LEGACY_CACHE is renamed CACHE_3_1.

test/lib/compliance-required-tools-applicability.test.js

The #3520 test that pinned 3.1.29 as partial now checks both sides of the boundary. On 3.1.23 every bundle stays partial. On 3.1.29 the unclaimed protocol's bundles are not_applicable, while the claimed protocol keeps its controller gap.

.changeset/capability-rollup-3-1-24.md

Patch changeset for @adcp/sdk.

What this unblocks for 3.1 sellers

Once hosted grading pins a release with this change:

• The declined, non-compliance_testing scenarios of the media-buy protocol bundle stop counting: audience, clicks and reach flows, frequency caps, inventory lists, proposals, refine, vendor-metric optimization, creative-library and governance scenarios.
• The universal wholesale_feed_* bundles stop counting for a seller that does not offer wholesale feeds.

The media-buy bundle can still stay partial through scenarios whose gate also involves compliance_testing.*. #2960 refuses any predicate set that contains a compliance_testing path, even when the failed predicate is the product one.

• On 3.1.29 that is create_media_buy_async, gated on compliance_testing.scenarios containing force_create_media_buy_arm.
• From 3.1.31 it also includes canonical_formats, canonical_custom_formats, canonical_experimental_formats, measurement_accountability, vendor_metric_accountability and universal stale-response-advisory. These AND a product predicate (for example media_buy.buying_modes contains wholesale) with compliance_testing.scenarios.

Claimants still need the tools, the controller and passing assertions.

Tests

• storyboard-capability-rollup.test.js: 46/46 pass with no skips (3.0.25, 3.1.31, 3.2.8 and 3.3.0-beta.5 caches present).
  • The two changed tests fail on the original source.
  • They also fail with a mutation that applies the rule to every 3.1 patch.
• compliance-required-tools-applicability.test.js: the re-pinned test passes, and fails on the original source.
• Seven related files (rollup, required-tools applicability, bundle results, phase capability cascade, universal metadata, OAuth metadata graph, request-signing vector transport): 131/132, no skips. The one local failure also fails locally on the original source, and passes on main CI.
• tsc --noEmit, eslint (no new findings) and prettier --check pass.

Out of scope / follow-ups

• AND-composed compliance_testing.* gates over-cap a seller that fails the product predicate: Capability rollup: don't let an AND-composed compliance_testing.* predicate cap a scenario the agent declined on product grounds #3701.
• SDK 16.x keeps the 3.2 cutoff.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve.

Backports the AdCP 3.2 "Applicability order" capability-rollup rule to 3.1.24+. The reviewer verified carriesApplicabilityOrderRule is correctly shaped: compareAdcpVersionStrings parses major.minor only (comply.ts:1474-1483) so 3.1.x < 3.2, and the exact-patch regex ^3\.1\.(\d+)$ bounded at patch ≥ 24 is the sole 3.1-line path. Fails closed on absent/invalid/v-prefixed/prerelease versions, each pinned by test. Guard stays narrow (compliance_testing.* gates, unmet requirements, execution evidence still cap). Tests add positive+negative boundary coverage (3.1.23 partial / 3.1.24+ passing); runner.ts change is comment-only and accurate. Changeset present and patch is appropriate for a grading alignment with no wire/export impact — satisfies DR-0033 minor-release pathway and changeset-vs-wire-impact rules.

No blocking or Medium findings. No gated paths, no high-risk flag, no team gate. Falls through to row 9.