fix infer logic of agent_run_time from the attached input trainer

[rsareddy0329]

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Claude finished @rsareddy0329's task in 55s —— View job

---

Code Review

• Read PR diff (/tmp/pr.diff)
• Review source change in multi_turn_rl_evaluator.py
• Review added/updated tests

---

The change is focused and correct: _resolve_trainer_defaults now handles the nested shape that AgentRFTJob.agent_config actually returns ({"BedrockAgentCoreConfig": {"AgentRuntimeArn": ...}} / {"CustomAgentLambdaConfig": {"LambdaArn": ...}}) while preserving the old flat-key fallbacks. Customer-provided agent_config still wins (guarded by if self.agent_config is None), so it's backward compatible. Test coverage is good — both nested shapes, both flat fallbacks, and the "customer value not overwritten" case.

A couple of minor points, neither blocking:

1. Potential AttributeError if a config key is present but None.

bedrock_cfg = resolved_agent.get("BedrockAgentCoreConfig", {})
lambda_cfg = resolved_agent.get("CustomAgentLambdaConfig", {})

dict.get(key, {}) only returns the {} default when the key is absent. If the key is present with an explicit None value (e.g. {"BedrockAgentCoreConfig": None}), bedrock_cfg becomes None and the following bedrock_cfg.get(...) raises AttributeError. If the upstream job can emit that shape, prefer resolved_agent.get("BedrockAgentCoreConfig") or {} to coerce None to {}.

2. Unit test relies on type(trainer).__name__ = "MultiTurnRLTrainer" on a MagicMock.
This works because _resolve_trainer_defaults gates on the class name, and it's a nice way to exercise the parsing logic in isolation — no concern, just noting it's a somewhat brittle pattern if the gating check ever changes to isinstance.

Overall this looks good to merge. The core fix is sound and well-tested.

[github-actions]

Claude finished @rsareddy0329's task in 43s —— View job

---

Code Review

• Read PR diff (/tmp/pr.diff)
• Review source change in multi_turn_rl_evaluator.py
• Review added/updated tests

---

The core fix is sound: _resolve_trainer_defaults now unpacks the nested shape AgentRFTJob.agent_config actually returns ({"BedrockAgentCoreConfig": {"AgentRuntimeArn": ...}} / {"CustomAgentLambdaConfig": {"LambdaArn": ...}}) while keeping the flat-key fallbacks, and the if self.agent_config is None guard preserves customer-provided values. Coverage is good — both nested shapes, both flat fallbacks, and the "don't overwrite customer value" case.

One point from the prior review is still unaddressed in the current diff:

1. AttributeError when a config key is present but None (multi_turn_rl_evaluator.py:260-261).

bedrock_cfg = resolved_agent.get("BedrockAgentCoreConfig", {})
lambda_cfg = resolved_agent.get("CustomAgentLambdaConfig", {})

dict.get(key, {}) returns the {} default only when the key is absent. If the upstream job emits the key with an explicit None value — e.g. {"BedrockAgentCoreConfig": None, "CustomAgentLambdaConfig": {"LambdaArn": "..."}}, which is a very common serialization shape when one of two mutually-exclusive configs is unset — then bedrock_cfg is None and bedrock_cfg.get("AgentRuntimeArn") raises AttributeError. Recommend coercing to {}:

bedrock_cfg = resolved_agent.get("BedrockAgentCoreConfig") or {}
lambda_cfg = resolved_agent.get("CustomAgentLambdaConfig") or {}

This is cheap insurance and the more likely-to-occur shape given only one config is populated at a time. Whether it's blocking depends on whether the producer can emit None-valued keys — worth confirming, and adding a {"BedrockAgentCoreConfig": None, ...} unit case to lock it in.

2. (non-blocking) Unit test uses type(trainer).__name__ = "MultiTurnRLTrainer" on a MagicMock.
Fine as-is since the gating is a class-name string check, but brittle if that gate ever moves to isinstance. Just noting.

Otherwise this looks good — focused, backward compatible, and well tested.

Note: the inline-comment tool wasn't available in this run, so findings are consolidated here with file/line references.