feat: support injection_config to disable runtime state injection in benchmarks

- Add InjectionConfig passthrough in AsAgentWrapper.reply()
- Disable inject_runtime_state in BaseAgenticAnswerStep to avoid
  wall-clock time conflicting with benchmark query_time anchors
- Disable inject_runtime_state in beam/lme llm_judge calls

reme/components/agent_wrapper/as_agent_wrapper.py

@@ -11,7 +11,7 @@ from pathlib import Path
 from typing import Any, TYPE_CHECKING
 from uuid import uuid4
 
-from agentscope.agent import Agent, ContextConfig, ReActConfig
+from agentscope.agent import Agent, ContextConfig, InjectionConfig, ReActConfig
 from agentscope.agent._config import ModelConfig
 from agentscope.event import (
     DataBlockDeltaEvent,
@@ -337,6 +337,7 @@ class AsAgentWrapper(BaseAgentWrapper):
             model_config=ModelConfig(**(kwargs.get("model_config") or {})),
             context_config=ContextConfig(**(kwargs.get("context_config") or {})),
             react_config=ReActConfig(**(kwargs.get("react_config") or {})),
+            injection_config=InjectionConfig(**(kwargs.get("injection_config") or {})),
         )
 
         if isinstance(inputs, str):

reme/steps/benchmark/base/agentic_answer.py

@@ -54,6 +54,9 @@ class BaseAgenticAnswerStep(BaseStep):
             "system_prompt": sys_prompt,
             "job_tools": ["search", "add_draft", "read_all_draft"],
             "react_config": {"max_iters": self.MAX_ITERATION},
+            # Benchmarks anchor "now" via query_time in the system prompt; the
+            # real wall-clock time injected by AgentScope would conflict with it.
+            "injection_config": {"inject_runtime_state": False},
             "tool_context_id": tool_context_id,
         }
 

reme/steps/benchmark/beam/llm_judge.py

@@ -77,7 +77,11 @@ async def _llm_equivalence(agent_wrapper, reference: str, system: str) -> bool:
     )
     user_prompt = f"First snippet: {reference}\n\nSecond snippet: {system}"
 
-    result = await agent_wrapper.reply(user_prompt, system_prompt=system_prompt)
+    result = await agent_wrapper.reply(
+        user_prompt,
+        system_prompt=system_prompt,
+        injection_config={"inject_runtime_state": False},
+    )
     raw = (result.get("result") or "").strip().lower()
     return "yes" in raw
 
@@ -277,7 +281,7 @@ class BeamRubricJudgeStep(BaseStep):
         for item in rubric:
             prompt = judge_template.replace("<rubric_item>", item).replace("<llm_response>", llm_response)
 
-            result = await self.agent_wrapper.reply(prompt)
+            result = await self.agent_wrapper.reply(prompt, injection_config={"inject_runtime_state": False})
             raw = (result.get("result") or "").strip()
 
             try:

reme/steps/benchmark/lme/llm_judge.py

@@ -59,6 +59,7 @@ class LmeAnswerJudgeStep(BaseStep):
         result = await self.agent_wrapper.reply(
             user_prompt,
             system_prompt=self.prompt_format(judge_prompt_key),
+            injection_config={"inject_runtime_state": False},
         )
 
         raw_answer = (result.get("result") or "").strip()