[tinalenguyen]

Adds DecisionGuardrail, which runs DecisionModel checks at three points in a session. Each check is a decorated handler that gets the results and returns what should happen next.

guardrail = DecisionGuardrail(session, model="openai/gpt-6-luna-decisions")

@guardrail.on_user_turn(decisions={
    "abusive": Probability("The user is abusive or threatening."),
    "resolved": Probability("The caller's issue is resolved."),
})
def before_reply(ev: DecisionEvaluated) -> Block | Steer | None:
    if ev.value("abusive", 0.0) > 0.8:
        return Block(say="I'm going to end the call here.")
    if ev.value("resolved", 1.0) < 0.3:
        return Steer(instructions="The issue is not resolved. Ask what is still wrong.")
    return None

@guardrail.on_agent_sentence(decisions={"off_policy": Probability("...")}, on_error="block")
def during_reply(ev: DecisionEvaluated) -> Block | None: ...

@guardrail.on_agent_turn(decisions={"empathetic": Probability("...")})
def after_reply(ev: DecisionEvaluated) -> None: ...

Hooks

Hook	Where it runs	Can return
on_user_turn	After the agent's on_user_turn_completed, before the reply is scheduled	Block, Steer, None
on_agent_sentence	Between the LLM text and the TTS / transcript, per sentence	Block, None
on_agent_turn	On conversation_item_added for assistant messages	None

• on_user_turn
  • Eager evaluation (default): each final transcript re-evaluates the turn so far, cancelling the previous request, so the result is usually ready at end of turn. If the final transcript differs, it's evaluated again.
  • Block: no reply. The user's message stays in the history, and say is spoken if given.
  • Steer: adds a system message to this turn's context only. A preemptive reply is then discarded by the existing equivalence check and regenerated. With a realtime model, it is passed as the reply's instructions instead.
• on_agent_sentence
  • Splitting: sentences are split with retain_format=True, so the released text is exactly what the LLM produced. They're checked concurrently and released in order.
  • Block: cancels the LLM, so there's no more text and no tool calls after the blocked sentence. Sentences already allowed finish playing and are the only ones kept in the history. say is spoken after the reply, unless it was interrupted or never played.
• Errors: on_error ("allow" / "block") decides what a failed check means: the evaluation fails or times out, any decision comes back without a result, or the handler raises. The timeout covers the evaluation and an async handler together.
• Multiple results: the first Block wins; otherwise every Steer applies.
• One per session: a check can evaluate any mix of Probability, Choice and Score, and a hook can have several handlers, so one guardrail covers a session. It is exposed as session.guardrail; creating a second one raises.

Not in this PR

• Holding playout until a check passes, and a reference-counted playout gate shared with AMD.
• say() text and realtime-model replies don't go through on_agent_sentence.

Tests

• tests/test_guardrail.py: unit tests for results, errors, timeouts, eager reuse, and the sentence gate.
• tests/test_guardrail_session.py: through AgentSession with the fake models. Covers Block and Steer on a user turn, a sentence Block mid-reply, and on_agent_turn.

[devin-ai-integration]

Devin Review found 1 new potential issue.

1 flag not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)

[devin-ai-integration]

🟡 Agent history helper lacks required documentation

The new _agent_history method has no docstring. Repository rules require documentation for every new method.

Suggested change

	def _agent_history(self) -> llm.ChatContext:
	# the context the reply is generated from, before on_user_turn_completed edits it
	try:
	return self._session.current_agent.chat_ctx
	except RuntimeError:
	return self._session.history
	def _agent_history(self) -> llm.ChatContext:
	"""Return the active agent's chat context, or session history if inactive."""
	try:
	return self._session.current_agent.chat_ctx
	except RuntimeError:
	return self._session.history

---

Was this helpful? React with 👍 or 👎 to provide feedback.