Feat support private messages

[Bekaboo]

Closes #

📑 Description

Added private message (DM) support for multi-agent interaction.

Unlike private actions, which are only visible to the author and recipients, but not to others.
Agents can set the to field in AgentAction class to specify the recipients.

It is called private actions instead of "private messages" as the to field is not limited to the "speak" action.

Any action, e.g. "non-verbal communication", "action", or even "leave" can have such property.

This PR is based on #329.

Co-author @ruiyiw

✅ Checks

• My pull request adheres to the code style of this project
• My code requires changes to the documentation
• I have updated the documentation as required
• All the tests have passed
• Branch name follows type/descript (e.g. feature/add-llm-agents)
• Ready for code review

ℹ Additional Information

---

Note

Introduce private actions via AgentAction.to with validation and per-recipient observations; switch action space to string literals; update server, docs, examples, and tests.

• Core changes:
  • Messages: Add AgentAction.to for private recipients with validation; to_natural_language prefixes [private to ...].
  • Env (ParallelSotopiaEnv):
    • Per-agent observation filtering for private actions via _actions_to_natural_language_for_viewer.
    • Validate actions with context (agent_names, sender) in step/astep.
    • Replace Discrete with LiteralSpace for action_type (string literals).
  • Server: Validate agent actions’ recipients; on error, notify agent and retry once.
• Docs:
  • Describe private messages, visibility rules, and to validation in messages, envs/parallel, concepts/agents, and examples index.
  • Note action_type is now a string literal in action space.
• Examples:
  • Add examples/experimental/multi_agents_private_dm/ (README, scripts, JSON configs) demonstrating multi-agent DMs and JSON logging.
• Tests:
  • Add tests for private visibility and invalid recipients; adjust evaluator test names; minor benchmark test refactors.
• Misc:
  • Minor signature cleanup in LLMAgent.act.

^{Written by Cursor Bugbot for commit beffe0d. This will update automatically on new commits. Configure here.}

[cursor]

This PR is being reviewed by Cursor Bugbot

Details

You are on the Bugbot Free tier. On this plan, Bugbot will review limited PRs each billing cycle.

To receive Bugbot reviews on all of your PRs, visit the Cursor dashboard to activate Pro and start your 14-day free trial.

[cursor]

Bug: Retry validation failure crashes without error handling

When an agent generates an invalid action with bad recipients, the code retries once by asking the agent to regenerate. However, the second AgentAction.model_validate() call at line 196-202 is not wrapped in a try-except block. If the agent still produces invalid recipients after the retry, the validation error will propagate as an uncaught exception and crash the episode. This makes the system fragile when LLM agents occasionally fail to follow the messaging protocol correctly.

 

[cursor]

Bug: Extra space in to_natural_language output formatting

The to_natural_language method in AgentAction has an extra space in the return statement. The recipients_prefix on line 140 already includes a trailing space when to is set (f"[private to {self.to}] "), but line 155 adds another space (f"{recipients_prefix} {action_str}"). This results in double spaces for private messages (e.g., "[private to ...] said:"), and a leading space for public messages (e.g., " said:"). This formatting issue will cause the test assertions in test_parallel_sotopia_env_multi_agents_private_messages to fail since they expect strings like 'agent1 said: "..."' without extra spaces.