[rollout] fix: prompt2text decoding for SingleTurnAgentLoop

[jnash10]

Fixes #5276

What does this PR do?

Fixes prompt IDs not being decoded for SingleTurnAgentLoop. Issue was the single agent loop lacked the rollout_trace_op decorator.

Patch: Add Weave Tracing to SingleTurnAgentLoop

Summary

Add @rollout_trace_op decorator to SingleTurnAgentLoop.run() to enable Weave tracing with token2text support.

Problem

When using SingleTurnAgentLoop (the default agent), Weave traces only show AsyncLLMServerManager.generate calls. The token2text=True option doesn't add prompt_text and response_text fields to the trace output.

This happens because SingleTurnAgentLoop.run() lacks the @rollout_trace_op decorator that ToolAgentLoop.run() has.

Component	Has @rollout_trace_op?
ToolAgentLoop.run	Yes
SingleTurnAgentLoop.run	No (missing)
AsyncLLMServerManager.generate	Yes

Solution

Add the @rollout_trace_op decorator to SingleTurnAgentLoop.run().

File Changed

verl/experimental/agent_loop/single_turn_agent_loop.py

Diff

 from verl.experimental.agent_loop.agent_loop import AgentLoopBase, AgentLoopOutput, register
 from verl.tools.utils.tool_registry import initialize_tools_from_config
 from verl.utils.profiler import simple_timer
+from verl.utils.rollout_trace import rollout_trace_op

 logger = logging.getLogger(__file__)
 logger.setLevel(os.getenv("VERL_LOGGING_LEVEL", "WARN"))
@@ -40,6 +41,7 @@ class SingleTurnAgentLoop(AgentLoopBase):
         tool_list = initialize_tools_from_config(tool_config_path) if tool_config_path else []
         self.tool_schemas = [tool.tool_schema.model_dump(exclude_unset=True, exclude_none=True) for tool in tool_list]

+    @rollout_trace_op
     async def run(self, sampling_params: dict[str, Any], **kwargs) -> AgentLoopOutput:
         messages = list(kwargs["raw_prompt"])

Why This Works

SingleTurnAgentLoop.run() already returns AgentLoopOutput with the required fields:

• prompt_ids (line 71)
• response_ids (line 72)

The add_token2text function in rollout_trace.py checks for these exact fields. Once the method is decorated and traced, token2text=True correctly adds prompt_text and response_text to the output.

Result

After this patch:

• SingleTurnAgentLoop.run appears as parent traces in Weave (matching ToolAgentLoop behavior)
• token2text=True adds prompt_text and response_text fields
• AsyncLLMServerManager.generate appears as nested child traces

[gemini-code-assist]

Code Review

This pull request correctly enables Weave tracing for SingleTurnAgentLoop by adding the @rollout_trace_op decorator to its run method. This change aligns its behavior with ToolAgentLoop and fixes the issue where prompt and response text were not being decoded in traces when token2text=True. The implementation is straightforward and correct. I have reviewed the changes and have not identified any issues.

[wuxibin89]

Please format code: https://github.com/verl-project/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting