Open Multi-Agent Files

Execution routing and governance

• Added pluggable execution routing for runTeam() through explicit mode, custom ExecutionRouter implementations, and the built-in DeterministicRouter.
• Added structured governance declarations for required or preferred roles, ordered review paths, budget-aware degradation, and post-execution governanceConclusion checks.
• Added privacy-preserving execution receipts and routing-decision trace linkage.
• Consequential tools can now be declared with ToolDefinition.consequential. Undeclared runs expose a machine-readable disclosure flag and can require confirmation through onToolCall.
• Automatic routing now recognizes structured Chinese, Japanese, and Korean goals using script-aware information length.

Scheduling and task execution

• Task DAG execution is event-driven by default: downstream work starts as soon as its dependencies complete.
• Added configurable dependency-first, round-robin, least-busy, capability-match, and weighted composite scheduling strategies.
• Agents and tasks can declare structured capabilities and hard requirements, with optional strict assignee validation.
• TeamRunResult.taskResults preserves task-scoped results by task ID.
• Explicit tasks can pass raw, structured, or combined dependency payloads with bounded role and provenance metadata.
• Added OrchestratorConfig.onTaskDispatch(task) for native per-task pipeline approval.

Model routing and reliability

• Model routes can declare ordered fallbacks for retryable worker-provider failures.
• Fixed CJK keyword extraction and zero-score fallback so eligible agents are selected consistently.
• Governed planOnly runs now return the declared role DAG without executing agents.
• Execution modes, governance floors, and per-run budget conflicts now follow a documented precedence order and disclose overrides or degradation.

Compatibility

• Existing onApproval integrations retain legacy round scheduling and callback semantics.
• Progress events from independent DAG branches may now interleave. Consumers should correlate events by task ID instead of relying on round adjacency.
• Raw dependency output remains the default. Structured handoffs, governance declarations, consequential confirmation, and custom execution routing are opt-in.
• New result fields remain optional for compatibility with older serialized results and caller-authored fixtures.

Packages

• @open-multi-agent/core: 1.13.0
• create-oma-app: 0.6.0; generated starters pin core 1.13.0
• @open-multi-agent/otel: remains at 0.1.0 and is not republished

Install

:::bash
npm install @open-multi-agent/core@1.13.0
npm create oma-app@latest my-oma

Thanks to @LambIessz for the retryable model-route fallback in [#424].