@wfgraph/core@3.0.0

Major Changes

• #145 708ac73 Thanks @alandotcom! - Replace the persisted Group exitNodeId field with exitNodeIds, and support lookup groups whose exits share one downstream endpoint.

Minor Changes

• #128 c2ae17b Thanks @alandotcom! - Add a build agent to the workflow editor: a chat panel that reads the extension
  catalog and the open workflow, then edits the canvas as it answers.

  A host turns it on with a new agent option on createWfGraphApp, carrying an
  OpenAI API key and optionally a model id and an OpenAI-compatible endpoint.
  Absent or blank, the agent is off: no model is called and the editor shows no
  panel, so an adopter who wants no AI in their editor writes nothing.

  The agent has fourteen tools. Seven read (search the action catalog, describe one
  action's config and output fields, list Events and integrations, read the graph,
  list the template tokens a step may reference, and validate the workflow) and
  seven write (add, update and delete a step, connect and disconnect steps, declare
  the Lifecycle Rules, and write a Condition's test). They run on the server against
  the graph the editor sends with each turn, and the resulting graph streams back
  so the canvas redraws as the agent works. The edit lands through the editor's own
  save path, so it is one undo step and the autosave persists it.

  GET /api/extensions now answers agent: { enabled } beside catalog, which is
  how the editor knows whether to offer the panel. agent.chat is the first
  streaming RPC procedure in the contract, declared as an oRPC event iterator.

  The effect catalog moves to 4.0.0-rc.110, which the @effect/ai-openai
  provider names as its peer.

Patch Changes

• #151 1da1baf Thanks @alandotcom! - Improve build-agent authoring for Wait steps, Event Split outlets, condition reference fields, and draft publication blockers.

• #128 c2ae17b Thanks @alandotcom! - Prevent the build agent from applying workflow edits that violate graph topology rules.

• #148 01da8e0 Thanks @alandotcom! - Stop sending start and result payloads on the run-list procedures.

  getExecutions polls every two seconds while the Runs tab is open, and
  getExecutionsGlobal pages the dashboard. Neither list paints input or
  output, yet both selected those JSONB columns and redacted them on every
  answer. Payloads stay on getExecutionLogs, which is fetched for the one open
  run.