v5.15.0

Minor Changes

• 6c1d5d0: resolveAccount now receives auth context, checkGovernance parses MCP envelope shapes correctly, and the seller skill documents alternative transports.

  resolveAccount(ref, ctx) now receives { toolName, authInfo }. Adapters that front an upstream platform API (Snap, Meta, TikTok, retail media networks) need the caller's OAuth token to look up the upstream account. Previously authInfo was only available inside handlers, forcing resolvers to return a thin stub and re-resolve the platform account on every handler call. Single-arg resolvers (async (ref) => ...) remain valid — TypeScript allows a shorter parameter list.

  dispatchTestRequest(request, { authInfo }). Test harnesses can simulate the authInfo that serve({ authenticate }) would populate, so resolveAccount and handler tests cover auth-sensitive paths without spinning up HTTP. Never mount this behind an HTTP route — extras.authInfo bypasses authenticate.

  checkGovernance now extracts from structuredContent or content[0].text before falling back to top-level fields. Fixes a latent bug where the helper returned "missing required fields" when the governance agent responded with a conformant MCP CallToolResult envelope rather than spreading the payload at the root. The single-agent JSDoc example no longer references a fabricated ctx.account.governanceAgentUrl field — it now shows the real sync_governance → Account.governance_agents[] flow.

  Multi-agent governance helper is deferred pending spec resolution on adcontextprotocol/adcp#3010 — sync_governance allows up to 10 governance agents per account but the check_governance request and the protocol envelope only thread a single governance_context per lifecycle. The SDK will ship an aggregation helper once the spec picks an interpretation.

  Skill docs. skills/build-seller-agent/SKILL.md gains an Alternative Transports section covering the createAdcpServer().connect(transport) pattern — multi-host HTTP on a single process and stdio — for cases where serve()'s single-publicUrl-per-process model doesn't fit.

Patch Changes

• a050729: Two regressions from the 5.14 train (closes #862). Both restore documented
  behavior — no new surface, no new policy. 5.14.0 consumers should upgrade
  directly to this release; no code changes required.

  (1) Schema loader: flat-tree domain $ref resolution

  ensureCoreLoaded pre-registered only core/ and enums/ before AJV
  compile. Tool schemas in flat-tree domain directories — governance/,
  brand/, property/, collection/, content-standards/, account/,
  signals/ — ship alongside sibling building-block fragments they $ref,
  and those siblings were never registered. First compile of e.g.
  governance/sync-plans-request.json threw can't resolve reference /schemas/3.0.0/governance/audience-constraints.json.

  The loader now walks every directory outside bundled/ and pre-registers
  non-tool JSON fragments — covering core/, enums/, pricing-options/,
  error-details/, extensions/, and every flat-tree domain's sibling
  building blocks. Tool request/response files stay lazy-compiled so
  relaxResponseRoot still applies to response variants.

  Blast radius is broader than storyboards. The same getValidator is
  wired into strict-mode request/response validation
  (AdcpClient({ strict: true }), createAdcpServer default validation,
  validateOutgoingRequest / validateIncomingResponse, the dispatcher
  middleware, and TaskExecutor). Any 5.14.0 server-side adopter running
  strict validation on governance/brand/property/signals/collection/
  content-standards/account tools was silently throwing on first call;
  those paths are fixed by this release too.

  (2) create_media_buy enricher: fixture-per-package precedence

  The fixture-authoritative refactor in 5.14 (#816) set every task's
  top-level merge to fixture-wins, but the nested-package merge in
  create_media_buy kept the prior builder-authoritative precedence.
  Storyboards that authored explicit product_id / pricing_option_id /
  bid_price on packages[0] had those values overridden by the first
  discovered product's pricing_options[0] — e.g. a seller's
  pinnacle_news_video_premium_pricing_0 replaced the fixture's
  cpm_guaranteed, failing create-buy with INVALID_REQUEST.

  Real seller ids in the fixture now win over discovery. Discovery
  still gap-fills when the author omits per-package ids — the
  behavior generic single-package storyboards rely on. Auction/CPM
  bid_price synthesis only fires when the fixture didn't author
  one, so bid-floor-boundary tests keep their explicit values.

  Sentinel placeholders pass through to discovery. The upstream
  universal compliance storyboards (adcontextprotocol/adcp:
  universal/deterministic-testing.yaml, error-compliance.yaml,
  idempotency.yaml, domains/media-buy/state-machine.yaml) ship
  packages[0] fixtures with product_id: "test-product" and
  pricing_option_id: "test-pricing" expecting the runner to
  substitute the seller's discovered identifiers. The enricher
  recognizes those two literals as sentinels and defers to discovery
  when either appears. Real seller ids (cpm_guaranteed,
  sports_display_auction, any non-sentinel string) keep winning.

  If your storyboard wants placeholder-then-discovery semantics for a
  new field, author $context.<key> substitution rather than a magic
  literal — the intent is explicit at the fixture level and the
  sentinel allowlist stays small.

  Out of scope

  Issue #862 also flagged activate_signal as "same pattern". The
  enricher is not FIXTURE_AWARE — the outer merge lets the storyboard's
  $context.first_signal_pricing_option_id overlay the enricher's pick,
  and both resolve from the same signals[0].pricing_options[0]. The
  mismatch reporters saw (po_prism_abandoner_cpm sent,
  po_prism_cart_cpm accepted) traces to seller catalog inconsistency
  between get_signals and activate_signal, not SDK synthesis.
  Follow-up in #870: have the storyboard runner emit a hint when a
  response's available: list excludes a context-derived value, so the
  reporter-facing symptom stops looking identical to an SDK bug.

• d856e1e: Triage routine now runs a mandatory pre-PR expert review on the diff (code-reviewer + domain expert in parallel) before opening the PR, capped at 2 review→fix iterations. Sign-offs recorded in the PR body.