v0.0.5

Added

• Plan-based coordination — the built-in understand coordination now outputs a structured plan (action: "plan") with typed steps instead of a reactive delegation loop. Each step has an id, optional agent, prompt, and dependsOn (string ids). Independent steps run in parallel; dependent steps resolve in order and receive their dependency outputs injected into the prompt. This eliminates infinite re-delegation loops and makes multi-agent execution deterministic, observable, and bounded to one plan round.
• "plan" event — fires on the conversation when the coordinator produces a plan, giving the application visibility into the execution strategy before it runs.
• "plan-step" event — fires started/completed/failed per step with the step's output text, for progress UIs and status tracking.
• "agent-delta" event — a delegated specialist's LLM output streamed in realtime, tagged with the agent name. Distinct from text-delta (the top-level reply): this is live progress for UX observation and is not coalesced into the conversation's reply object.
• finalizeAfterToolRound option — sub-generations skip the redundant answer round after tool execution, synthesizing output directly from tool results. Cuts wall time ~30%.
• agents action support — the flattened delegate tool still accepts the simpler action: "agents" / tasks syntax alongside plan. Both route through the same deterministic executor.
• Delegated sub-generation token cap — sub-generations within a plan step bound output to 512 tokens and limit tool iterations to 3 (was 10).
• retryDirectAnswer option — when a multi-agent coordination answers directly without planning, the framework retries once.

Changed

• understand prompt tightened — encourages the model to speak the full answer in its narration BEFORE emitting the tool call, eliminating the separate composition LLM round-trip.
• delegate tool description — updated to emphasize planning as the primary action.

Fixed

• Sub-generation text deltas no longer pollute the conversation stream — specialist output now streams through agent-delta instead of text-delta, so the top-level reply object (coalesced by ConversationStream) is not fragmented by intermediary agent narration.