v5.19.0

Minor Changes

• af944a1: feat(AgentClient): add AgentClient.fromMCPClient() factory for in-process MCP transport

  Adds a new static factory method that accepts a pre-connected @modelcontextprotocol/sdk Client instance instead of a URL-based agent config. This enables compliance test fleets to wire up a full AgentClient against an InMemoryTransport pair without an HTTP loopback server.

  MCP only. This factory wraps an MCP Client from @modelcontextprotocol/sdk. There is no equivalent in-process bridge for A2A today — for A2A agents, run them on a loopback HTTP server and use the standard AgentClient constructor with the agent's agent_uri.

  Key behaviors preserved over the in-process path:

  • adcp_major_version is injected on every tool call
  • idempotency_key is auto-generated for mutating tasks
  • isError envelopes surface as TaskResult<{ success: false }>
  • HTTP-only methods (resolveCanonicalUrl, getWebhookUrl, registerWebhook, unregisterWebhook) throw descriptive in-process guard errors
  • Endpoint discovery and SSRF validation are bypassed for the sentinel URI

  Exports the new InProcessAgentClientConfig type for typed factory usage.

• efbe785: Add pgBackend.probe() and serve({ readinessCheck }) for fail-fast pool validation

  Sellers wiring createIdempotencyStore({ backend: pgBackend(pool) }) from a DATABASE_URL env var previously got a silent failure mode: a bad URL (typo, deprovisioned DB, missing creds) lets the server boot successfully, advertise IdempotencySupported, then fail every mutating call indefinitely.

  This release adds:

  • pgBackend.probe() — runs SELECT 1 FROM "<table>" LIMIT 0 at startup, validating both connectivity and that the idempotency table has been migrated. Throws a descriptive error naming the table, root cause, and remediation steps.
  • IdempotencyStore.probe() — delegates to backend.probe() when the backend implements it; no-ops for memoryBackend.
  • probeIdempotencyStore(store) — convenience export for callers that manage their own lifecycle (Lambda, custom HTTP frameworks).
  • ServeOptions.readinessCheck?: () => Promise<void> — called before httpServer.listen(). The server never accepts connections if the check throws, so a misconfigured pool crashes the process at deploy time rather than silently failing live traffic.

  Wire the probe in serve():

  const store = createIdempotencyStore({ backend: pgBackend(pool), ttlSeconds: 86400 });
  pool.on('error', err => console.error('pg pool error', err)); // prevent crash on idle-client errors
  serve(createAgent, {
    readinessCheck: () => store.probe(),
  });

  readinessCheck is general-purpose — use it for any startup dependency check, not just idempotency.

  Non-breaking. createIdempotencyStore remains synchronous. Existing callers require no changes. Option A (async constructor) is tracked separately as a future major-version enhancement.

• a26db16: Storyboard runner: add $generate:opaque_id substitution and context_outputs[generate] for threading runner-minted task IDs through multi-step lifecycle storyboards.

  $generate:opaque_id and $generate:opaque_id#<alias> work identically to $generate:uuid_v4 / $generate:uuid_v4#<alias> but carry explicit task-ID semantics. Both share the same alias cache namespace.

  context_outputs entries now accept generate: "opaque_id" | "uuid_v4" as an alternative to path:. When generate is set the runner mints (or reuses, via alias-cache coherence) a UUID at post-step time and writes it into $context.<key> for subsequent steps. If an inline $generate:opaque_id#<key> substitution already ran in the same step's sample_request, the generator reuses that value — the two forms are alias-coherent.

  ContextProvenanceEntry.source_kind and ContextValueRejectedHint.source_kind gain a 'generator' variant for accurate diagnostic attribution. ContextOutput.path is now optional (mutually exclusive with the new generate field).

Patch Changes

• c58ff99: Fix get_media_buys convention extractor poisoning context during multi-page pagination walks (#998). The extractor unconditionally captured media_buys[0].media_buy_id from every successful get_media_buys response. When a storyboard walks multi-page results, the page-1 response carries pagination.has_more: true — buys[0] is not the canonical buy, it is just the first item in a list slice. The captured ID was then picked up by the request-builder enricher on step 2 and injected as media_buy_ids: [that_id], turning the pagination continuation into a single-ID lookup. The agent returned one buy with has_more: false, total_count: 1, failing total_count: 3 storyboard assertions.

  The extractor now skips extraction when pagination.has_more === true, matching the conservative === true convention used elsewhere in the codebase (hasMorePages() in validations.ts). When has_more is absent or false — i.e., a terminal or single-page response — extraction proceeds as before. This unblocks get-media-buys-pagination-integrity in adcontextprotocol/adcp#3122 from upgrading to the seeded multi-page walk model used by list_creatives and other paginated storyboards.