[tool, rollout] fix: handle optional interaction kwargs

[denismegerle]

What does this PR do?

Improves robustness of multi-turn tool agent loops when using an interaction config: samples without extra_info.interaction_kwargs are treated as having no interaction, while misconfigured interaction_kwargs.name values fail fast.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/verl-project/verl/pulls?q=is%3Apr+ToolAgentLoop+interaction_kwargs
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

• pre-commit run --all-files --show-diff-on-failure --color=always
• Added CPU unit test: tests/experimental/agent_loop/test_tool_agent_loop_interaction_kwargs_on_cpu.py

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# When interaction_config_path is set, rows can omit interaction configuration:
extra_info = {}

# Rows that want interaction should provide a valid name:
extra_info = {"interaction_kwargs": {"name": "weather"}}

# If a name is provided but not configured, the loop now raises ValueError.

Design & Code Changes

• Make interaction_kwargs optional when an interaction config is present.
• Only enter INTERACTING / process assistant messages when an interaction is actually initialized.
• Raise on unknown interaction_kwargs.name to avoid silently skipping misconfigurations.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• (not required) Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: Covered by new CPU unit test.
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• (not related) If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Denis Megerle seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[gemini-code-assist]

Code Review

The pull request correctly addresses the handling of optional interaction configurations in the tool agent loops. By guarding the interaction initialization and state transitions, it allows for datasets with mixed interaction requirements. The addition of a fail-fast check for unknown interaction names improves error reporting. I have suggested a minor refinement to ensure the code is robust against missing or null extra_info fields.

[gemini-code-assist]

Accessing kwargs["extra_info"] directly can raise a KeyError if the key is missing, or an AttributeError if extra_info is None. Since this PR aims to improve robustness for optional configurations, it is safer to handle these cases explicitly to avoid crashes on minimal or null input data.

Suggested change

	interaction_kwargs = kwargs["extra_info"].get("interaction_kwargs", None) or {}
	interaction_name = interaction_kwargs.get("name", None)
	extra_info = kwargs.get("extra_info") or {}
	interaction_kwargs = extra_info.get("interaction_kwargs") or {}
	interaction_name = interaction_kwargs.get("name")

[gemini-code-assist]

Accessing kwargs["extra_info"] directly can raise a KeyError if the key is missing, or an AttributeError if extra_info is None. It is recommended to use a safer extraction method to ensure the loop handles samples without any extra info gracefully.

Suggested change

	interaction_kwargs = kwargs["extra_info"].get("interaction_kwargs", None) or {}
	interaction_name = interaction_kwargs.get("name", None)
	extra_info = kwargs.get("extra_info") or {}
	interaction_kwargs = extra_info.get("interaction_kwargs") or {}
	interaction_name = interaction_kwargs.get("name")