This pull request adds result. agentToolInvocation for nested Agent.asTool() runs and keeps result.runContext focused on plain execution context. The main runtime change lives under packages/agents-core/src/{agent,result,runContext,runState}.ts, with supporting tests in packages/agents-core/test/**, and it also exports CompletedAgentToolRunResult so customOutputExtractor can be annotated explicitly by SDK users.

The change is intentionally simpler than the earlier runContext-union approach: agent-tool invocation metadata now lives on the result object that consumes it, RunContext no longer carries agent-tool-specific shape, and serialized RunState writes top-level agentToolInvocation metadata.

Minimal usage example:

const billingTool = billingAgent.asTool({
  toolName: 'billing_agent',
  customOutputExtractor(result: CompletedAgentToolRunResult<any, typeof billingAgent>) {
    console.log(result.agentToolInvocation.toolName);
    console.log(result.agentToolInvocation.toolCallId);
    return String(result.finalOutput ?? '');
  },
});

Behaviorally, this gives users a stable public way to inspect nested agent-tool calls without reading result.state._context, and it keeps resumed nested runs consistent with non-resumed runs.