endInvocation

Requests the current LLM agent to stop after the current step completes.

Scope is exactly this LLM agent: the per-step loop in LlmAgent.executeTurns exits after the current step, and this agent's remaining after-agent callbacks (the checks at the end of BaseAgent.runAsync) are skipped.

The flag does NOT propagate to any other agent. Enclosing workflow agents (SequentialAgent, LoopAgent, ParallelAgent) do not read InvocationContext.isEndOfInvocation, and each child agent runs under its own context copy produced by InvocationContext.forAgent(...) / branching, so the mutation never reaches the parent's context. In Sequential[A, B], a callback in A calling endInvocation() still lets B run. This matches Python ADK (sequential_agent.py:91-99, loop_agent.py:113-122, per-agent context copy in base_agent.py:433) and Java ADK (LoopAgent.java:146 ->takeUntil(hasEscalateAction), per-agent toBuilder() in InvocationContext.java:270). To break out of a LoopAgent, set EventActions.escalate = true instead.

Mirrors Python ADK's callback_context._invocation_context.end_invocation = True and Java ADK's EventActions.setEndInvocation(true) / setEndOfAgent(true).