feat(governance): add seller policy and change-right contracts - #6794
feat(governance): add seller policy and change-right contracts#6794bokelley wants to merge 26 commits into
Conversation
45e5a03 to
924893a
Compare
924893a to
da00e20
Compare
|
Reviewed at Two small items:
Supportive of landing this. |
…matchers may carry paths Buyers reconciling a disclosed rejection against their any_of[] matcher list should compare full matcher entries, not origins — an agent_url matcher may include a path and therefore be narrower than the disclosed origin, causing a false match on origin-only comparison. Suggested by @lukasz-pubx in review of #6794.
|
Thanks for the thorough review, @lukasz-pubx. Item 2 —
Item 1 — decision records + Generated by Claude Code |
![aao-secretariat[bot]](https://avatars.githubusercontent.com/u/221690821?s=60&v=4){kind=small}
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated paths touched without required approval.
This PR modifies protocol source under static/schemas/source/** (and .agents/**), which are under a hard, deterministic approval gate (gated_paths: true). The current review_decision is REVIEW_REQUIRED, not APPROVED, so the gate is NOT satisfied and the PR cannot be auto-approved (decision table row 2).
The reviewer findings are clean: the change relaxes error.code from required to optional on the newly-added, unreleased reported-outcome-error.json schema — a non-breaking loosening on an unreleased surface — and it is propagated coherently across schema, docs (report_plan_outcome.mdx), the governance handler, and both server and .cjs tests, with a minor changeset covering the additive surface. No critical/high/medium findings were reported.
Escalation reasons:
- Touches gated protocol-source files (e.g.
static/schemas/source/governance/reported-outcome-error.json,static/schemas/source/enums/error-code.json, and many others understatic/schemas/source/**, plus.agents/sdk-shim-ledger.json) that require human/CODEOWNERS approval;review_decisionisREVIEW_REQUIRED.
Once a human/CODEOWNERS approval sets review_decision to APPROVED, this PR falls through to a normal clean approve on the current findings.
Why human review
- Gated paths touched (static/schemas/source/** and .agents/**) — hard approval gate not satisfied: review_decision is REVIEW_REQUIRED, not APPROVED. Human/CODEOWNERS review required.
- This PR touches a path under a hard, non-overridable approval gate (.agents/sdk-shim-ledger.json (modified) matches
.agents/**; static/schemas/source/account/sync-governance-response.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-media-buy-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-available-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/product-allowed-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/sla-window.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/x-entity-types.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/canonical-media-buy-action-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/error-code.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/media-buy-action-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/error-details/governance-agent-not-accepted.json (added) matchesstatic/schemas/source/**; static/schemas/source/error-details/policy-violation.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/accepted-governance-agents.json (added) matchesstatic/schemas/source/**; static/schemas/source/governance/get-plan-audit-logs-response.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/policy-category-definition.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/policy-entry.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/report-plan-outcome-request.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/reported-outcome-error.json (added) matchesstatic/schemas/source/**; static/schemas/source/index.json (modified) matchesstatic/schemas/source/**; static/schemas/source/media-buy/acceptance-context.json (added) matchesstatic/schemas/source/**; static/schemas/source/media-
…(truncated)…
aao-secretariat
Bot
added
the
ladon/needs-human-review
|
Issue #6823 proposes a new Generated by Claude Code |
|
WG ratification is now recorded for all four RFCs:
The decision records are committed in 6060673, formal decision comments are posted on each RFC, the |
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated paths touched without required review approval.
This PR modifies files under hard approval gates (static/schemas/source/** and .agents/**) — the protocol wire-shape source of truth. gated_paths is true and review_decision is REVIEW_REQUIRED (not APPROVED), so decision-table row 2 forces escalation regardless of how clean the diff is.
The reviewer found no critical/high/medium findings this pass. The delta since the prior escalate adds only four well-formed governance decision records (DR-0015–DR-0018), whose factual claims check out against the schema/docs/SDK changes. This is a net-clean iteration, but the hard gate cannot be auto-satisfied by Ladon.
Additionally, this PR introduces new protocol schema fields, new error codes (GOVERNANCE_AGENT_NOT_ACCEPTED, GOVERNANCE_UNAVAILABLE), new media-buy acceptance-policy schemas, and new enum values — breaking/normative-class protocol surface changes whose ratification is an explicit human act per the constitution's breaking-class hard gate. Human/CODEOWNERS review is required before merge.
Escalation reasons
- Gated paths touched (
static/schemas/source/**,.agents/sdk-shim-ledger.json) — hard approval gate not satisfied:review_decisionisREVIEW_REQUIRED, notAPPROVED. - Protocol wire-shape changes (new schemas, fields, error codes, enum values) require human ratification per the breaking-class governance gate.
Why human review
- Gated paths touched (static/schemas/source/** and .agents/sdk-shim-ledger.json) — hard approval gate not satisfied: review_decision is REVIEW_REQUIRED, not APPROVED. Human/CODEOWNERS approval required.
- Protocol wire-shape surface changed (new acceptance-policy schemas, new error codes, new enum values, new required/optional fields) — breaking-class ratification is a human act per constitution; never auto-approved absent APPROVED review.
- This PR touches a path under a hard, non-overridable approval gate (.agents/sdk-shim-ledger.json (modified) matches
.agents/**; static/schemas/source/account/sync-governance-response.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-media-buy-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-available-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/product-allowed-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/sla-window.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/x-entity-types.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/canonical-media-buy-action-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/error-code.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/media-buy-action-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/error-details/governance-agent-not-accepted.json (added) matchesstatic/schemas/source/**; static/schemas/source/error-details/policy-violation.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/accepted-governance-agents.json (added) matchesstatic/schemas/source/**; static/schemas/source/governance/get-plan-audit-logs-response.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/policy-category-definition.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/policy-entry.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/report-plan-outcome-request
…(truncated)…
|
Addressed the #6823 documentation note in 27ff9e1. The new sell-side governance boundaries page makes the spend-approval asymmetry explicit and separates buyer authority, seller acceptance, negotiated change rights, and seller-internal review. It also documents the seller-disposition/audit and governance-agent dependency boundaries introduced by this PR. The PR now carries |
…icies # Conflicts: # docs.json # server/src/training-agent/task-handlers.ts # server/src/training-agent/tenants/router.ts # server/src/training-agent/tenants/tenant-smoke.test.ts # tests/docs-nav-validation.test.cjs
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
KonstantinMirin
left a comment
KonstantinMirin
left a comment
There was a problem hiding this comment.
Review — PR #6794
Overview — This PR closes #6749/#6750/#6757/#6758/#6823, and it picked the right shape for all four RFCs. Facets landed on the generic policy-category-definition.json with a category-interpolating validator. change_terms went inside commercial_terms, so the existing terms_digest covers it. #6757 re-scopes three existing error codes and adds one optional origin rather than a new code. #6758's matcher union is a real discriminated any_of, with unresolvable verification routed to GOVERNANCE_UNAVAILABLE. The problem is reachability, not design: every surface this PR adds is gated on >= 3.2-beta.7, and the reference seller's highest negotiable release is 3.2-beta.6, so none of it is reachable on any transport — and the diff moved two previously wire-level tests off the wire to compensate. Three of #6749/#6750's own test-shaped acceptance criteria are unmet, and the page that closes #6823 has no inbound link. The design is sound; what follows is making it reachable and provable.
Should fix
1. The wire gate cannot open, and the proofs moved off the wire instead of the wire moving.
supportsSellerGovernanceDiscovery() returns true only for >= 3.2-beta.7 (server/src/training-agent/types.ts:64, a hardcoded prerelease >= 7). The highest release in SUPPORTED_RELEASE_VERSIONS is 3.2-beta.6 (task-handlers.ts:2633, tenants/router.ts:135), and resolveServedAdcpVersion rejects an unlisted prerelease outright — no caller can open the gate. That one predicate gates media_buy.acceptance_policy_discovery, adcp.governance_enforcement.accepted_governance_agents, and the sync_governance rejection itself (account-handlers.ts:1594). grep -rn '3\.2-beta\.7' server/src returns exactly one match: SELLER_GOVERNANCE_DISCOVERY_ADCP_VERSION at types.ts:29, which nothing reads. The threshold is written twice — once as a named constant, once as a magic 7 inside a character-for-character clone of supportsGetProductsRejected — so the two can only agree by coincidence, and they already disagree with the servable set.
The consequences are visible in the diff. Both new storyboards gate on capabilities the seller never emits, so the runner skips them not_applicable, which run-storyboards.ts:932 counts as clean. training-agent.test.ts converted two tests from simulateCallTool(server, 'get_adcp_capabilities', …) to direct handleGetAdcpCapabilities({}, { …, servedAdcpVersion: '3.2-beta.7' }) calls, and the sole assertion of #6758's credential-non-echo MUST calls handleSyncGovernance() in-process (account-handlers.test.ts:1073) while every sibling sync_governance test in that file goes through simulateCallTool. Meanwhile tenant-smoke.test.ts asserts on the real HTTP route that the capability blocks are undefined and that binding https://untrusted-governance.example/mcp returns status: 'synced' with the endpoint persisted. Those wire tests are correct — they document that the feature is off.
Root: a wire-gated feature with two writers for its checkpoint, and a checkpoint the server does not serve. Nothing in the build can observe that a gated capability is unreachable, because an unsatisfiable requires_capability is indistinguishable from one that is legitimately not applicable.
Fix: add 3.2-beta.7 to the release list and advance CURRENT_ADCP_VERSION / TRAINING_AGENT_CURRENT_ADCP_VERSION to it (the immutability of the published 3.2-beta.6 bundle is why a new checkpoint is needed, not a reason to ship an unreachable one) — or lower the gate to >= 6 if the checkpoint genuinely cannot move here. Either way, express both predicates as one-liners over a single atLeastAdcpVersion(served, minimum) comparator driven by the exported *_ADCP_VERSION constants, reusing the parseAdcpReleaseVersion / compareAdcpReleaseVersions pair this PR already extended at task-handlers.ts:2673. Then revert the two converted tests to simulateCallTool and drop the servedAdcpVersion: '3.2-beta.7' injections. If the checkpoint cannot move, the two scenarios must not be listed in requires_scenarios yet — registering a scenario whose gate can never open is a false green.
2. ~300 lines of new change-term enforcement sit in a branch that excludes exactly the state they handle, and the dead copy is fail-open.
task-handlers.ts:18479-18787 is guarded by !(acceptedCommercialTerms && Object.hasOwn(acceptedCommercialTerms, 'change_terms')) and then reads acceptedTerms.change_terms off the same object at :18485:
const acceptedCommercialTerms = mediaBuy?.acceptedProposal?.commercial_terms as unknown as Record<string, unknown> | undefined;
if (mediaBuy?.acceptedProposal && args.revision === mediaBuy.revision
&& !(acceptedCommercialTerms && Object.hasOwn(acceptedCommercialTerms, 'change_terms'))) {
const acceptedTerms = mediaBuy.acceptedProposal.commercial_terms as unknown as Record<string, unknown>;
const acceptedChangeTerms = Array.isArray(acceptedTerms.change_terms) ? … : [];So acceptedChangeTerms is always [], changeTermByAction is always empty, permitsChange() always returns false, and budgetConstraintViolation() always returns undefined at its first guard. Every new call site (:18565, :18619, :18723, and the permitsChange guards at :18577, :18630, :18738, :18768) lies inside that branch. The REQUOTE_REQUIRED behavior the storyboard grades comes from the other implementation — rejectUnavailableAction → enforceAcceptedChangeConstraint, invoked earlier in the same function at :18471.
The two copies have already diverged in the unsafe direction. enforceBudgetConstraint returns cannot_preflight when the current value is missing or a money constraint has a mismatched currency (:4384, :4348). The inline copy drops the bound to undefined — that is, allows — in both cases (:18502, :18506-18510). One rule, two evaluators, one of them permissive on unknown.
Root: the mutation path carries two parallel authorization systems. The correct one is already wired for control_media_buy; the PR copied the new rule inward and then added a guard that made the copy dead.
Fix: delete acceptedChangeTerms / changeTermByAction / permitsChange / budgetConstraintViolation and the permitsChange guards from handleControlMediaBuyUnlocked, leaving that block as the legacy no-terms envelope check it is guarded to be. Anything genuinely wanted from it becomes a case in enforceBudgetConstraint, with a control_media_buy step in change_rights_state_projection.yaml that reaches it.
3. The seller advertises the acceptance contract but implements no evaluator, so #6749's preflight loop is exercised by nobody.
acceptance_context: true is advertised at both capability sites (task-handlers.ts:15222, tenants/router.ts:852) and read nowhere — grep -rn acceptance_context server/src returns those two literals and one test type declaration, zero reads. The schema says that flag exists precisely to distinguish a seller that interprets the context "from a legacy seller that does not interpret it", and docs/protocol/get_adcp_capabilities.mdx:446 says true "confirms the seller evaluates the buyer's structured … facts". The reference seller makes exactly the claim the flag exists to deny it.
The conformance artifacts grade the declaration, not the evaluation. acceptance_policy_discovery.yaml:118 sends subject_facets: ["federal_election"], which is not a facet of political_advertising or of any other category (candidate_or_party, election_or_ballot_measure, issue_advocacy, legislation_or_regulation, government_information, political_merchandise, incidental_or_news_reference), so it matches no rule in the profile the very next assertion discovers. Nothing catches it — x-entity: governance_policy_category_facet is only pattern-checked. The scenario's only product-side assertion is that a product echoes back its own seeded acceptance_policy_profile_ids.
Two of #6749's acceptance criteria fall out of this. "The schema can also encode one non-political conditional category" has no artifact. subject_category is political_advertising at all 13 occurrences across tests/, static/registry/, and the storyboards, and static/registry/policy-categories/pharmaceutical_advertising.json has no facets. "Conformance tests cover overlapping rules and the precedence of product-specific policy over seller defaults" has no artifact either. The restrictive-composition rule is normative in docs/media-buy/media-buys/policy-compliance.mdx, and no step and no test exercises it. Invert the precedence rule in any implementation and the whole suite stays green.
Fix: implement the resolver over the already-published catalog — it is a pure function of profile plus context — and grade that in acceptance_policy_discovery.yaml: a context selecting a prohibited rule must project differently from one selecting conditional, and a context matching no published rule must resolve to unknown, never allowed. Correct the facet to a registry value that actually selects a rule. Add facets to one non-political category with a conditional profile case in tests/seller-governance-policy-surfaces.test.cjs. Advertise and enforce are one decision: until the seller evaluates the context, do not advertise the flag — an absent capability already means unknown, which is the fail-closed reading the schema prescribes.
4. One capability contract, two writers — and the author built the right pattern one property away.
task-handlers.ts:15218 and tenants/router.ts:848 each hand-assemble the same four-field acceptance_policy_discovery object from the same three imported constants, with different gates: the handler adds ctx.tenantId === 'sales' || ctx.tenantId == null, the router has no null-tenant arm. The sibling value added in the same hunks does it right — TRAINING_ACCEPTED_GOVERNANCE_AGENTS (account-handlers.ts:48) is one exported constant consumed at both sites. The grading consequence is concrete: the byte-digest recomputation test (training-agent.test.ts:15563-15588) covers only the handler copy, while the router copy — the path that serves the sales tenant over HTTP — is asserted only negatively (tenant-smoke.test.ts:611, :634: toBeUndefined()).
The same shape repeats twice more in this diff. router.ts:135 keeps a byte-identical private copy of SUPPORTED_RELEASE_VERSIONS, so adding '3.1' took four synchronized edits. router.ts:441-471 hand-maintains the "is this scenario handled in-process" chain that comply-test-controller.ts:1205-1221 and :1370-1374 already compute; the two lists have drifted, and registering one scenario in this PR took four edits across two layers.
Root: projectTenantCapabilities re-parses an already-serialized MCP response buffer (router.ts:705-733) and rewrites capability content. That inversion predates this PR, but this PR made it a second author of a new contract instead of leaving it a pure selector. The missing abstraction is a capability model owned by the business layer; today the model exists only as literals inside two serializers.
Fix: export builders with declared return interfaces next to TRAINING_ACCEPTED_GOVERNANCE_AGENTS — acceptancePolicyDiscoveryCapability(servedVersion, tenantId) and governanceEnforcementCapability(tasks, servedVersion) — each owning its own version and tenant gate, so both writers reduce to one spread and a future schema field is a compile error at both sites. Export the release list and import it in the router. Export the controller's local-scenario predicate. Extend the existing byte-digest assertion to the router projection so the served path is the pinned one.
5. static/schemas/source/** is the declared single writer of every wire contract, but nothing downstream derives from it — so each contract is re-authored by hand in TypeScript.
One disease, four surfaces:
- Types.
change-term.jsonis a closed object andchange-term-constraints.jsona discriminated union, yet the reference seller re-implements both asRecord<string, unknown>+isRecord()probes acrossderiveAvailableActionsFromAcceptedChangeTerms,enforceAcceptedChangeConstraint,enforceBudgetConstraint, andproposalChangeTermsForPurchases. "Does this buy have negotiated change terms" is hand-rolled four times (:3936,:4247,:18478,:18486), each re-deriving the absent/non-array/present tri-state the fail-closed semantics depend on.MediaBuyProductAllowedActionState.constraintsisRecord<string, unknown>(types.ts:568); the action-mode union is spelled inline four times and duplicated across two value-identical schema enums, so addingseller_managedtook six parallel edits. One regression makes this a defect rather than a style note.rejectUnavailableAction's return type was widened from{ errors: TaskError[]; context?: unknown } | nulltoRecord<string, unknown> | null(:4246), while all four sibling builders in the same diff keep the concrete type. It is then invoked forcontrol_media_buywithargs as unknown as UpdateMediaBuyArgs(:18471) — a cast the type system cannot check, and one that is false in general. - Tool schema and runtime validator.
governance/reported-outcome-error.jsonis transcribed by hand intoREPORTED_OUTCOME_ERROR_INPUT_SCHEMA(governance-handlers.ts:44-101) and re-implemented a third time imperatively incloneReportedOutcomeError(:107-232). The MCP copy's internal$refs are absolute into the embedding document (#/properties/error/definitions/…), so it only resolves while spliced at exactlyproperties.error. The 16 KiB byte cap exists only in the imperative copy, so the published contract never states the bound the seller enforces. The new test compilesGOVERNANCE_TOOLS' owninputSchemaand validates it against itself. - Repository and API model.
deserializePublishedPolicy(row: any)(policies-db.ts:109-144) reconstructs the publicPolicyfield-by-field out of untyped JSONB with?? nulldefaults and synthesizes domain state the row does not carry (source_type: 'registry',review_status: 'approved').acceptance_profileisRecord<string, unknown>/z.record(...)(policies-db.ts:20,registry.ts:1039) although its shape is fully declared by this same PR inmedia-buy/acceptance-policy-profile.jsonand validated key-by-key bycheck-registry-completeness.cjs— andPolicySchemafeeds the published OpenAPI, so consumers getadditionalProperties: truefor a closed, digest-pinned shape.issuer, added in the same hunk, gets a concrete type; the profile does not. - Field-name lists.
COMPACT_PRODUCT_FIELDS(task-handlers.ts:6231) is a third hand-kept copy of the product-field vocabulary and namesacceptance_policy_refs, a field that exists in no schema, no doc, no registry, and no test.
Root: no derivation step runs from the generated schema artifacts into the reference seller — not for types, not for MCP tool schemas, not for the database and API model. Hand-mirroring is therefore the path of least resistance, and this PR added four concepts at once. "The SDK snapshot lags the schema" is the reason, not a justification: types.ts already hosts MediaBuyAvailableActionState for exactly this purpose. Widening to Record<string, unknown> is the same missing-contract smell one notch up — it still gives callers no key or value checking.
Fix, one surface at a time:
- Declare
ChangeTerm,ChangeTermConstraints(discriminated onkind),ChangeTermServiceModeandMEDIA_BUY_ACTION_MODESintypes.ts, mirroring the schemas. Parse once at the trust boundary (parseChangeTerms(commercialTerms): ChangeTerm[]), let every downstream function take the typed value, and delete the per-field probes. - Restore the concrete rejection type (
type MediaBuyActionRejection = { errors: TaskError[]; context?: unknown }) and pass a domainAttemptedMediaBuyActionEntry[]instead of the cast. - Load the generated
reported-outcome-error.jsononce and splice it underdefinitionswith self-relative pointers. Validatereq.errorwith the Ajv-compiled schema on all transports. LeavecloneReportedOutcomeErrorresponsible only for what JSON Schema cannot express — byte cap, prototype and accessor hygiene, and safe deep clone — and state the byte bound in the schemadescription. - Derive
Policyfrom the Zod schema, parsecanonical_contentwith a declaredCanonicalPolicyDocumentso a malformed publication fails loudly, and typeacceptance_profilewith the profile shape. - Delete
acceptance_policy_refsand deriveCOMPACT_PRODUCT_FIELDSfrom the canonical-product schema's property names. - Have
canonical-media-buy-action-mode.json$ref/allOf-narrow the base enum, so a member cannot land in one file and not the other.
6. Assertions that cannot fail, and proofs that never run.
Five instances, one mechanism:
change_rights_state_projection.yaml:217writes the expected code undercode:where every othercheck: error_codeentry in the corpus usesvalue:— 96value:against exactly onecode:, the one this PR adds, on #6750's fail-closed branch. The error-code lint and the contradictions lint both harvest onlyvalue/allowed_values, so the step asserts nothing andACTION_NOT_ALLOWEDis never checked against the canonical enum. Both siblings in the same file usevalue:, and this step also omits theirexpect_error: true/negative_path:.negative_path: wrong_status(:314) andconstraint_exceeded(:350) are outside the documentedschema_invalid | payload_well_formedenum.lint-storyboard-sample-request-schema.cjs:344-348treats anything that is notpayload_well_formedas skip-schema-validation, so the static shape check is off for both steps — and both send deliberately well-formed payloads rejected on state/bounds grounds, which is thepayload_well_formedcase.scripts/run-storyboards-matrix.shis edited by this diff, butmin_clean/min_passedand theREQUIRED_CLEAN_/REQUIRED_EXACT_CURRENT_SALESlists are untouched while four scenarios joinrequires_scenarios.git diff main...HEAD -- scripts/run-storyboards-matrix.sh | grep -E 'min_passed|REQUIRED_'returns nothing. The repo already has the right instrument —compact_product_lifecycle:9:0is pinned exactly — and the new scenarios were not enrolled.tests/seller-governance-policy-surfaces.test.cjs(400 lines, 8 tests) is the proof artifact for all four RFCs' schema semantics and is referenced by nothing:grep -rn seller-governance-policy-surfacesover the head returns zero hits outside the file. Sibling.cjssuites run only becausepackage.jsonnames them, and the one glob runner excludes**/*.test.cjs. Deleteoriginfrompolicy-violation.jsonand no pipeline notices.- Two guards pass on their own output.
policy-publication-resolution.test.tsassertsresolvePolicyreturned the values its own stub supplied, plus two SQL-substring checks that pass on a semantically wrong query — delete the fail-closedif (version && policy.version !== version) return nulland both tests stay green. Andcheck-registry-completeness.cjs:174has no negative case: makeprofileContentDigest()returnprofile.content_digestand the suite stays green, which turns every stored digest self-attesting.
Root: every guard in the proof pipeline is a hand-maintained list or an unvalidated key — suites are opt-in by editing package.json, scenarios are pinned by editing a shell array, matcher keys and negative_path values are unchecked vocabulary. The default state of a new proof is dormant, and a typo degrades an assertion to a no-op rather than failing loudly.
Fix the five instances, then close the class that produced them.
Repair the instances:
- Change
code:tovalue:on that step and addexpect_errorandnegative_pathalongside it. - Set
negative_path: payload_well_formedon both out-of-enum steps. - Register the
.cjssuite in atest:script. - Enroll all four scenarios in
REQUIRED_EXACT_CURRENT_SALESwith their exactP/Stuples. The exact form catches a degradation into skips; required-clean does not. Raise thesalesandgovernancemin_passedfloors by the steps the scenarios contribute, and mirror both edits in.github/workflows/training-agent-storyboards.ymlper the keep-in-sync header. - Add the behavioral assertions to the resolver test: no publication row plus a current row at a different version resolves to
null, and a digest or canonical mismatch is not authoritative. Assert the parameter array rather than the SQL text. - Add the missing negative case for the profile-digest comparison.
Close the class:
- Extend
lint-storyboard-check-enum.cjsto reject a validation object that carries no recognized matcher key, and to validatenegative_pathagainst its declared enum. - Add a guard asserting that every
tests/*.test.cjsfile appears in sometest:*script.
The sibling available_actions.yaml family carries the same out-of-enum negative_path values, which is why the fix belongs in the lint.
7. Every safety branch is graded off the wire or not at all — the storyboards grade the accepted path.
- #6750's "an omitted action is not a negotiated right" has no scenario.
grep -rn not_supported_on_buy static/compliance/source/returns nothing; the branch exists attask-handlers.ts:4276-4280and is pinned only bycomply-test-controller.test.ts:1001. The one scenario assertingnot_supported_on_productis the pre-existingavailable_actions.yaml, which grades the product-template path, not the accepted-terms path. #6750's acceptance criteria name not-negotiated explicitly. - #6757's bounded-evidence family — over-byte, over-depth, over-wide
details, over-long property name, invalidrecovery, invalidclassification_source, and "the ledger is not mutated" — is asserted entirely by directhandleReportPlanOutcome(...)calls. Nothing proves a buyer on the wire receivesVALIDATION_ERRORand thatoutcomes_reportedstays0. - #6758's credential/URL non-echo MUST is asserted by one in-process
handleSyncGovernancecall (see finding 1). The storyboard asserts onlyfield_absent: accounts[0].governance_agents, never thathttps://untrusted-governance.example/mcpand the submitted credential are absent from the response. - #6757's one new normative field is graded nowhere on the wire:
failed_outcome_audit_persistence.yamlreportsdetails: { disposition: "declined", disclosure: "opaque" }— two keys that appear in no schema — and asserts ondisclosure.grep -rn origin static/compliance/source/protocols/governance/returns nothing. An implementation can pass conformance while never emitting the field #6757 was filed to add. - The disclosed arm of
governance-agent-not-accepted.jsonis produced by nothing:grep -rn attempted_agent_origin server/src static/compliancereturns no matches. The recovery path the error code's ownsuggestioninstructs buyers to follow — "select a governance agent matchingerror.details.accepted_governance_agents" — exists only as an Ajv shape assertion.
Fix by adding the missing steps to the scenarios that already exist:
- Add one
update_targetingattempt againstchange_rights_state_buyassertingACTION_NOT_ALLOWEDwitherrors[0].details.reason == "not_supported_on_buy". The fixture already carries a closed five-entrychange_termslist. - Add one over-bound
report_plan_outcomestep withexpect_error: trueandvalue: VALIDATION_ERROR, followed by aget_plan_audit_logsstep asserting that the ledger is unchanged. - Assert on the rejection step that the submitted URL host and credential appear nowhere in the response, and route the non-echo unit test through
simulateCallTool. - Report
details: { origin: "seller", seller_policy_ref: … }and assertentries[*].error.details.origin. Keep a second outcome carrying the bare{ origin: "seller" }so the scenario grades the genuinely opaque form too, and drop the inventeddispositionkey. - Have the seller return the disclosed arm for its already-public allowlist and grade both arms, asserting that
attempted_agent_originequals the stripped origin and not the submitted URL.
8. RFC 8785 canonicalization was forked, and the fork both mints and verifies immutable digests.
scripts/check-registry-completeness.cjs:32-63 adds a from-scratch JCS canonicalize / digest / policyContentDigest / profileContentDigest. Those functions mint every content_digest baked into the immutable publication migration (generate-policy-publication-migration.cjs:22,124 → 553_publish_political_acceptance_policies.sql) and every self-referential digest in static/registry/policies/*.json. The only test of those digests imports the same functions (tests/seller-governance-policy-surfaces.test.cjs:7, :115), so it proves the artifact matches its own producer and nothing about RFC 8785. The runtime and the documented consumer obligation use a different implementation — canonicalize from @adcp/sdk (plan-hash.ts:15, task-handlers.ts proposalTermsDigest), and registry.ts:1041 tells consumers to "canonicalize this object with RFC 8785 before verifying content_digest".
Root: canonicalization is a contract primitive with no shared home, so the validation layer grew its own — and that fork was promoted to the authority for a content-addressed, immutable-by-trigger registry. Publications are immutable, so a divergence in number formatting, escape selection, or key ordering ships as a permanently unverifiable published digest.
Reconsider the port itself before choosing how to test it. @adcp/sdk already implements RFC 8785, and the runtime and the documented consumer obligation both use it. Porting a second copy into the validation layer buys nothing the dependency does not already provide, and it costs a permanent maintenance obligation: every JCS edge case — number formatting, escape selection, key ordering, lone surrogates — becomes this repo's to track and to keep in step with the implementation buyers actually run. A spec-conformance primitive is the worst candidate for a hand-port, because the cost of a subtle divergence lands on immutable published digests that no later commit can correct.
Fix: use the existing implementation. Put one JCS module in one place, import it from the scripts, the server, and the tests, and back it with the canonicalize dependency — a thin CJS wrapper if the script boundary needs one. Then delete the fork.
Keep the local copy only if the CJS/ESM boundary genuinely blocks a direct import, and treat that as a temporary state with a tracked removal, not a resting place. In that case, add a cross-check asserting that digest(x) equals the SDK-canonicalized SHA-256 over a divergence-prone fixture set (non-ASCII, surrogate pairs, nested key ordering, integers against floats, -0, deep nesting), plus every checked-in static/registry/policies/*.json. training-agent.test.ts:15577-15581 already does this correctly for the catalog digest by recomputing from the served bytes, and is the pattern to copy.
9. New concepts and normative rules are restated per surface instead of declared once and referenced.
change_term_idplus the deprecatedterms_refblock (with itsdeprecated/x-deprecated-in/x-removed-inand 130-character description) is pasted verbatim into all threeoneOfbranches ofcore/canonical-media-buy-action.json(lines 26/27, 41/42, 56/57) and once more inmedia-buy-available-action.json. A 4.0 removal needs four identical edits.acceptance_policy_profile_idsis defined inline on bothcore/product.json:512andcore/canonical-product.json:77, and the descriptions have already diverged — only one states "the most restrictive matching disposition wins", so a reader of the compact product gets a weaker contract. The field name is also added to two separate selectable-field enums.reported-outcome-error.json:45re-declarescodeatmaxLength: 128where the canonicalcore/error.jsonsays 64, so the audit copy can validly retain a code the wire could never have carried.- The rule #6750 exists to establish — accepted terms are the ceiling — is stated in four places with two modal strengths:
proposal-negotiation.mdx:470says MUST NOT,update_media_buy.mdx:61says "cannot", pluscontrol_media_buy.mdxandproduct-discovery/index.mdx:209. A conformance reader cannot tell which is the requirement. docs/governance/sell-side-governance.mdx, the page that closes #6823, is unreachable. Thedocs.jsonnav entry added in 27ff9e1 no longer exists at the head (an upstream restructure to versioneddist/docs/...routes landed in between), andgrep -rn sell-side-governanceover the tree returns zero hits.docs-nav-validation.test.cjspasses, so the build is fine — the issue closes with a page nothing links to.GOVERNANCE_AGENT_NOT_ACCEPTED'senumDescriptionsentry (enums/error-code.json:187) is the only governance-family description with zero backticked symbols —GOVERNANCE_UNAVAILABLEhas 12,GOVERNANCE_DENIED54,PROVENANCE_VERIFIER_NOT_ACCEPTED24 — so its accordion indocs/snippets/compliance-error-codes.mdxrenderssync_governance,error.details, andGOVERNANCE_UNAVAILABLEas plain prose while every neighbour renders them as code.
Root: this repo's standard is that each new concept is its own $id'd schema $ref'd from every surface that carries it, and each normative rule has one home the task pages link to (as the sibling pages already do for url-canonicalization.mdx). This PR inlined per surface, which forks the contract at authoring time — and the fork has already produced a weaker contract for compact-product readers and a bound that contradicts core/error.json.
Fix each restatement at its source:
- Extract the shared action fields into an
$id'd definition and compose it per branch withallOf, so the discriminator survives. - Give both product schemas one
$ref'dacceptance-policy-profile-idsfield schema, and haveget-products-request.json'sfieldsenum composeproduct-fields.jsonrather than restate it. - Set
codetomaxLength: 64, or state in the description why the audit copy is deliberately wider. - Demote three of the four prose statements to pointers at the one normative home.
- Cross-link the governance page from
docs/governance/overview.mdx,docs/governance/campaign/specification.mdx, anddocs/accounts/tasks/sync_governance.mdx. - Put the symbols in the error-code description and its
errorRecoverysuggestion in code font.
10. The v6 facade writes a false account.sandbox to steer an internal partition, and the falsified ref outlives the request.
server/src/training-agent/v6-sales-platform.ts:355-376 rewrites an incoming account.sandbox: true to false for static: principals so that sessionKeyFromArgs (state.ts:718-723, which keys on canonical.sandbox === false) takes its shared-partition branch. The comment concedes the value is untrue. But the rewritten AccountRef is what the business layer receives, persists onto media-buy and creative state (task-handlers.ts:578, :12625, :13626) and compares with accountRefsOverlap (:13721) — an authority-scoping comparison. The stored account disposition depends on which entry point wrote it.
Root: the session-partition key is expressed as a mutated wire model instead of a derived key. An adapter layer is authoring a domain fact.
Fix: leave the AccountRef untouched and move the decision into key derivation — sessionKeyFromArgs, or a partitionHintFor(principal, ref) it consumes, treats public static: principals as the shared open partition explicitly, taking the hint as a parameter. The facade passes a hint; it never writes a false field value into a model other layers trust.
11. The new shim-ledger entry has no working reverse gear, and the mutations it authorizes fail open.
.agents/sdk-shim-ledger.json:43-59 takes the reach-in ledger from 2 entries to 3, authorizing a string patch against compiled vendor JS anchored on the whitespace-exact literal ' "cancellation_terms"\n]);' (scripts/overlay-compliance-cache.sh). lint-sdk-shims.cjs states the bar: shims should be "visible, owned, tied to an upstream ask, and removed when the SDK grows a public API". This entry fails two halves of that.
The entry's upstream is adcp-client#2664. Verified with gh issue view: #2664, #2665, adcp-client-python#1067, and #1069 are all OPEN and unassigned, so nobody is obligated to do the work that deletes the entry. And #2664's body ("feat(media-buy): unified product, proposal, and live-buy action assessment") never mentions the negotiation verifier, dist/lib/negotiation/verification.js, or its hand-maintained commercial_terms allowlist — closing #2664 in full does not satisfy the stated removalCondition. That is a permanent entry wearing a "status": "temporary" label.
Separately, the two new addLiteralToConstUnion(...) mutations at scripts/overlay-compliance-cache.sh:210-211 reuse a helper that returns silently when its anchors move (if (declarationStart < 0) return;), so a future SDK bump turns the ledgered shim into a no-op and the failure resurfaces as an opaque storyboard validation error. The same diff shows the correct shape 25 lines later: patchProposalVerificationFile re-checks its post-condition and throws.
Fix the ownership and the failure mode together:
- Self-assign #2664, #2665, adcp-client-python#1067, and #1069, so the parking spots are owned.
- Either add the concrete item to #2664 — "derive the negotiation verifier's
commercial_termsfield allowlist from the generated schema so additive optional fields need no consumer patch" — or file a dedicated issue and re-pointupstreamandremovalConditionat it. - Convert
addLiteralToConstUnionandaddLiteralToFirstActionUnionfromreturntothrowwhen the const, the union terminator, or theafterValueanchor is missing, so every ledgered mutation proves that it applied.
12. One documented URL-canonicalization rule, two implementations.
docs/reference/url-canonicalization.mdx gains a row (added by this diff) asserting that governance_agents[].url vs accepted_governance_agents.any_of[].agent_url is governed by the same eight-step algorithm as the preceding provenance-verifier row. In the reference seller the two rows run different code with observably different output. The new path calls canonicalTargetUri from @adcp/sdk/signing (account-handlers.ts:1575). The accepted_verifiers allowlist still uses the local canonicalizeAgentUrl (task-handlers.ts:4824), which lowercases scheme and host and strips a trailing slash, but strips no default ports and normalizes no dot-segments — despite its own doc comment claiming it does. An allowlist entry that matches for a creative verifier can fail to match for a governance agent.
The same constant also normalizes twice: TRAINING_ACCEPTED_GOVERNANCE_AGENTS advertises TRAINING_ACCEPTED_GOVERNANCE_AGENT_URLS verbatim (account-handlers.ts:49) while enforcement builds its set as TRAINING_ACCEPTED_GOVERNANCE_AGENT_URLS.map(canonicalTargetUri) (:1589), rebuilt on every iteration of the per-account loop. Today's three literals survive canonicalization unchanged, so nothing is broken yet — add a fourth with a default port, an uppercase host or a trailing slash and the seller advertises one string and enforces another. The constant's own doc comment calls itself "one source of truth for what is advertised and what is enforced"; it is single only at the string level.
Fix: route both allowlist comparisons through one canonicalizer implementing the documented algorithm, with the https/userinfo rejection layered on top for the governance path, and delete canonicalizeAgentUrl. Canonicalize the allowlist once at module scope and derive both the advertised matchers and the enforcement set from that one array.
Notes / prior-review follow-ups
- The prior rounds all landed: DR-0015…DR-0018 are committed with
Closesreferences (@lukasz-pubx item 1), thesync_governance.mdxmatcher-vs-origin clarifier is in place (item 2), anderror.codeoptionality is restored with both schema and runtime regressions (@bokelley's thread onreport-plan-outcome-request.json:192). None of those is re-raised. - Out of scope:
projectTenantCapabilitiesre-parsing an already-serialized MCP response buffer (tenants/router.ts:705-733) is a pre-existing inverted boundary. Reworking it is separate architectural work outside this diff — finding 4 asks only that the contract this PR adds there not acquire a second author. - Out of scope:
run-storyboards.tsdispatches all ~170 storyboards over MCP only, and there is no A2A/REST/ADK storyboard runner in the tree. That is a pre-existing property of the harness shared by every scenario, so it is not charged here — but it does mean none of the four new scenarios prove cross-transport serialization. - Out of scope: the real root of the compiled-JS overlay is upstream —
@adcp/sdk's negotiation verifier keeps a hand-maintainedcommercial_termsfield allowlist duplicating knowledge the generated schema already carries, so every additive optional field needs a monkey-patch. That fix lives in another repository; only the in-repo half is asked for in finding 11.
|
As author of #6757/#6758 (now DR-0017/DR-0018), a prioritization note on the review above: four of the findings determine whether those two ratified contracts are demonstrated by this PR rather than merely present, and I've verified each at
No position from me on the remaining implementation findings — though the drifted duplicate budget evaluator (allows on currency mismatch where the canonical helper fails closed) and the second RFC 8785 implementation minting immutable digests look like the two with real blast radius. None of this changes my support for landing the PR once the gates above make the new surfaces observable. |
|
All four prioritization points map to open review threads posted today at
Generated by Claude Code |
|
Addressed the full review set in cc9061d and resolved all 32 review threads. For the four prioritized proof gaps:
The broader review is also incorporated: duplicated capability/version/budget/digest logic was consolidated; proposal terms fail closed as the authority ceiling; opaque conditions, compound actions, aggregate package rights, and seller-managed paths were hardened; policy publications are append-only and digest-verifiable; registry/profile/schema composition is pinned and linted; SDK follow-ups are assigned; and the sell-side governance boundary documentation is linked from the main governance surfaces. Verification completed locally: full build/static worker, 4/4 canonical server-unit shards (6,810 tests), schema suite, JSON docs, migrations, OpenAPI, registry completeness, and both current and 3.0 storyboard matrices. The four new exact scenarios execute with zero skips: change rights 8P, acceptance discovery 3P, governance-agent binding 4P, and failed-outcome audit persistence 4P. |
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — hard approval gate not satisfied.
This PR touches files under gated paths (static/schemas/source/**, .agents/sdk-shim-ledger.json, .github/workflows/**) and review_decision is REVIEW_REQUIRED, not APPROVED. Per decision rule row 2, the outcome must be escalate (no critical/high findings exist to make this request-changes).
Additionally, this is a breaking-class governance surface change per repo context: new protocol schemas (acceptance-policy family), new error codes (GOVERNANCE_AGENT_NOT_ACCEPTED), new enum values, and new discriminated oneOf arms. Even though the reviewer found the surface uniformly additive and correctly discriminated with a correct minor changeset, protocol wire-shape ratification on this class of change is a human act per the constitution and is never auto-approved absent an APPROVED review.
Escalation reasons
- Gated paths touched (
static/schemas/source/**,.agents/sdk-shim-ledger.json,.github/workflows/training-agent-storyboards.yml) — hard approval gate;review_decisionis REVIEW_REQUIRED, not APPROVED. Human/CODEOWNERS approval required. - Protocol wire-shape surface expanded (new acceptance-policy schemas, new error code, new enum values, new optional fields, new oneOf arms) — governance/breaking-class ratification requires explicit human sign-off.
No Critical/High/Medium code findings were reported; migrations 557/558 are additive/transactional, schema↔docs coherence holds, and terms_ref is deprecated-with-compat rather than removed. The escalation is driven purely by the unsatisfied approval gate — once a human/CODEOWNERS approval is recorded, this can fall through to a normal decision.
Why human review
- Gated paths touched (static/schemas/source/**, .agents/sdk-shim-ledger.json, .github/workflows/training-agent-storyboards.yml) — hard approval gate not satisfied: review_decision is REVIEW_REQUIRED, not APPROVED. Human/CODEOWNERS approval required.
- Breaking-class protocol surface change (new acceptance-policy schemas, new GOVERNANCE_AGENT_NOT_ACCEPTED error code, new enum values, new optional fields, new discriminated oneOf arms) — ratification is a human act per the constitution; never auto-approved absent an APPROVED review.
- This PR touches a path under a hard, non-overridable approval gate (.agents/sdk-shim-ledger.json (modified) matches
.agents/**; .github/workflows/training-agent-storyboards.yml (modified) matches.github/workflows/**; static/schemas/source/account/sync-governance-response.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/acceptance-policy-profile-ids.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-media-buy-action-fields.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-media-buy-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-available-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-change-term-id.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-legacy-terms-ref.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/product-allowed-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/sla-window.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/x-entity-types.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/canonical-media-buy-action-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/enums/error-code.j
…(truncated)…
|
Final CI follow-up: merged current main, fixed the tenant capability schema assertion, compacted redundant enum types in MCP prompt views to restore the 400 KiB budget, and aligned storyboard floor tests with the ratcheted 571/161 step floors. The replacement run is fully green (all checks successful or neutral), the branch is clean/mergeable, Ladon added no new findings, and all review threads are resolved. The only remaining GitHub block is the required human approval. |
|
Final merge/CI follow-up at 76f5901:
The only remaining failed check is external reviewer infrastructure: Ladon completed with no buffered inline comments, then failed because it did not create its findings JSON. This reproduced on two reruns and has no repository finding to address. GitHub also still reports REVIEW_REQUIRED, so a human/CODEOWNERS approval is needed to record the WG approval and clear the hard gate. |
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — hard approval gate not satisfied and Breaking-class governance change requires human ratification.
No blocking findings — the reviewer's subsequent-pass review of the delta since f58408d surfaced no critical/high/medium findings. The protocol-surface changes were verified as coherent: the new condition_unresolved reason value aligns across action-not-allowed-reason.json, error-code.json, control-media-buy-request.json, and the corresponding docs/compliance scenario (no schema↔docs drift); the minor changeset is correct for additive enum/fields retaining 3.1 compatibility; policies-db.ts async conversions and fail-closed resolvePolicy are correct; the acceptance_profile requirements oneOf is discriminated by a kind const. This is a clean diff.
Why escalate, not approve:
- Gated paths (row 2):
gated_paths: true(touchesstatic/schemas/source/**,.agents/**,.github/workflows/**) andreview_decision: REVIEW_REQUIRED— NOTAPPROVED. The hard, non-overridable approval gate is not satisfied. Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded. - Breaking-class governance surface: New acceptance-policy schemas, new
GOVERNANCE_AGENT_NOT_ACCEPTEDerror code, new enum values, new discriminated oneOf arms, and new decision records DR-0015..0018. Per the constitution's Breaking-class hard gate, ratification is a human act and is never auto-approved absent an APPROVED review. - Prior decision was also
escalatefor the same gate reasons; the gate condition persists unchanged.
Escalation reasons:
- Gated files under
static/schemas/source/**,.agents/sdk-shim-ledger.json, and.github/workflows/training-agent-storyboards.ymlrequire human/CODEOWNERS approval;review_decisionisREVIEW_REQUIRED. - Breaking-class / Normative governance protocol change (new schemas, error code, enum values, DR-0015..0018) — ratification is a human act.
Why human review
- Gated paths touched (static/schemas/source/**, .agents/sdk-shim-ledger.json, .github/workflows/training-agent-storyboards.yml) — hard approval gate not satisfied: review_decision is REVIEW_REQUIRED, not APPROVED. Human/CODEOWNERS approval required.
- Breaking-class governance protocol surface change (new acceptance-policy schemas, new GOVERNANCE_AGENT_NOT_ACCEPTED error code, new enum values, new discriminated oneOf arms, DR-0015..0018) — ratification is a human act per the constitution; never auto-approved absent an APPROVED review.
- This PR touches a path under a hard, non-overridable approval gate (.agents/sdk-shim-ledger.json (modified) matches
.agents/**; .github/workflows/training-agent-storyboards.yml (modified) matches.github/workflows/**; static/schemas/source/account/sync-governance-response.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/acceptance-policy-profile-ids.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-media-buy-action-fields.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-media-buy-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/canonical-product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-available-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-change-term-id.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/media-buy-legacy-terms-ref.json (added) matchesstatic/schemas/source/**; static/schemas/source/core/product-allowed-action.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/product.json (modified) matchesstatic/schemas/source/**; static/schemas/source/core/sla-window.json (modified) matches `static/schemas/so
…(truncated)…
3 participants
Summary
Security and compatibility
3.1); the released3.1.19artifact remains the compatibility validatorValidation
SDK follow-ups
Closes #6749
Closes #6750
Closes #6757
Closes #6758
Closes #6823