feat(agent): add verification configuration for agents and update related components

backend/agents/create_agent_info.py

@@ -6,7 +6,7 @@
 
 from jinja2 import Template, StrictUndefined
 from nexent.core.utils.observer import MessageObserver
-from nexent.core.agents.agent_model import AgentRunInfo, ModelConfig, AgentConfig, ToolConfig, ExternalA2AAgentConfig, AgentHistory
+from nexent.core.agents.agent_model import AgentRunInfo, ModelConfig, AgentConfig, ToolConfig, ExternalA2AAgentConfig, AgentHistory, AgentVerificationConfig
 from nexent.core.agents.agent_context import ContextManagerConfig
 from nexent.memory.memory_service import search_memory_in_levels
 
@@ -484,6 +484,7 @@ async def create_agent_config(
         external_a2a_agents=external_a2a_agents,
         context_manager_config=cm_config,
         context_components=context_components,
+        verification_config=AgentVerificationConfig.model_validate(agent_info.get("verification_config") or {}),
     )
     return agent_config
 

backend/consts/model.py

@@ -2,7 +2,7 @@
 from typing import Optional, Any, List, Dict
 
 from pydantic import BaseModel, Field, EmailStr, ConfigDict, field_validator
-from nexent.core.agents.agent_model import ToolConfig
+from nexent.core.agents.agent_model import AgentVerificationConfig, ToolConfig
 
 from consts.prompt_template import PROMPT_GENERATE_TEMPLATE_FIELD_ALIAS_MAP
 
@@ -454,8 +454,16 @@ class AgentInfoRequest(BaseModel):
     group_ids: Optional[List[int]] = None
     ingroup_permission: Optional[str] = None
     enable_context_manager: Optional[bool] = None
+    verification_config: Optional[Dict[str, Any]] = None
     version_no: int = 0
 
+    @field_validator("verification_config", mode="before")
+    @classmethod
+    def normalize_verification_config(cls, value):
+        if value is None:
+            return None
+        return AgentVerificationConfig.model_validate(value).model_dump()
+
 
 class AgentIDRequest(BaseModel):
     agent_id: int
@@ -527,6 +535,7 @@ class ExportAndImportAgentInfo(BaseModel):
     author: Optional[str] = None
     max_steps: int
     provide_run_summary: bool
+    verification_config: Optional[Dict[str, Any]] = None
     duty_prompt: Optional[str] = None
     constraint_prompt: Optional[str] = None
     few_shots_prompt: Optional[str] = None

backend/database/agent_db.py

@@ -163,6 +163,7 @@ def create_agent(agent_info, tenant_id: str, user_id: str):
     """
     info_with_metadata = dict(agent_info)
     info_with_metadata.setdefault("max_steps", 15)
+    info_with_metadata.setdefault("verification_config", None)
     info_with_metadata.update({
         "tenant_id": tenant_id,
         "version_no": 0,  # Default to draft version
@@ -201,6 +202,7 @@ def create_agent(agent_info, tenant_id: str, user_id: str):
             "group_ids": new_agent.group_ids,
             "is_new": new_agent.is_new,
             "enable_context_manager": new_agent.enable_context_manager,
+            "verification_config": new_agent.verification_config,
             "current_version_no": new_agent.current_version_no,
             "version_no": new_agent.version_no,
             "created_by": new_agent.created_by,

backend/database/db_models.py

@@ -324,6 +324,7 @@ class AgentInfo(TableBase):
     current_version_no = Column(Integer, nullable=True, doc="Current published version number. NULL means no version published yet")
     ingroup_permission = Column(String(30), doc="In-group permission: EDIT, READ_ONLY, PRIVATE")
     enable_context_manager = Column(Boolean, default=False, doc="Whether to enable context management (compression) for this agent")
+    verification_config = Column(JSONB, doc="Layered ReAct self-verification configuration")
 
 
 class PromptTemplate(TableBase):

backend/prompts/managed_system_prompt_template_en.yaml

@@ -66,6 +66,11 @@ system_prompt: |-
      - Note that executed code is not visible to users. If users need to see the code, use '<DISPLAY:language_type>code</DISPLAY>' for displaying code.
      - **IMPORTANT**: After code execution, the system will return content with "Observation:" marker (this is the real execution result). Please continue your next thinking based on these real results. **Do NOT fabricate observation results before code execution.**
 
+  3. Self-verification:
+     - After critical events (tool calls, retrieval results, code execution, and final-answer preparation), the system may run explicit verification.
+     - If verification reports errors, insufficient evidence, incomplete parameters, or unreliable results, you must repair the issue, gather more evidence, call tools again, or clearly state what cannot be completed.
+     - The final answer is shown to the user only after verification passes. If the system returns Verification feedback, treat it as a real observation and continue revising.
+
   After thinking, when you believe you can answer the user's question, you can generate a final answer directly to the user without generating code and stop the loop.
 
   When generating the final answer, you need to follow these specifications:
@@ -178,3 +183,13 @@ final_answer:
     Original task: {{task}}
 
     Please provide a clear and concise summary of the work completed so far.
+
+
+verification:
+  pre_messages: |-
+    You are a strict verifier for a ReAct agent. Judge reliability only from the task, candidate answer, tool outputs, and observations. Do not output hidden chain-of-thought.
+    You must output JSON only.
+
+  post_messages: |-
+    Verify whether the candidate answer covers the user's intent, is grounded in observations, handles tool errors, uses trustworthy citations, and is formatted for users.
+    Output fields: passed, score, status, failed_criteria, checks, revision_instruction, user_visible_note.

backend/prompts/managed_system_prompt_template_zh.yaml

@@ -130,6 +130,11 @@ system_prompt: |-
      - 注意运行的代码不会被用户看到，所以如果用户需要看到代码，你需要使用'<DISPLAY:语言类型>代码</DISPLAY>'表达展示代码。
      - **重要**：代码执行后，系统会返回 "Observation:" 标记的内容（这是真实的执行结果）。请基于这些真实结果继续下一步思考，**不要在代码执行前自行编造观察结果**。
 
+  3. 自验证：
+     - 关键事件（工具调用、检索结果、代码执行、准备最终回答）后，系统会进行显式自验证。
+     - 如果自验证提示存在错误、证据不足、参数不完整或结果不可靠，必须优先修正、补充证据、重新调用工具，或清晰说明无法完成的部分。
+     - 最终回答只有在自验证通过后才会展示给用户；如果系统返回 Verification feedback，请把它视为真实观察结果继续修正，不要忽略。
+
   在思考结束后，当你认为可以回答用户问题，那么可以不生成代码，直接生成最终回答给到用户并停止循环。
 
   生成最终回答时，你需要遵循以下规范：
@@ -271,3 +276,13 @@ final_answer:
     原始任务：{{task}}
 
     请对迄今为止完成的工作进行清晰、简洁的总结。
+
+
+verification:
+  pre_messages: |-
+    你是 ReAct 智能体的严格验证器。请仅根据任务、候选答案、工具输出和观察结果判断答案是否可靠，不要输出隐藏思维链。
+    你必须只输出 JSON。
+
+  post_messages: |-
+    请验证候选答案是否覆盖用户意图、是否有观察结果支撑、是否处理了工具错误、引用是否可信、格式是否适合展示。
+    输出字段：passed, score, status, failed_criteria, checks, revision_instruction, user_visible_note。

backend/prompts/manager_system_prompt_template_en.yaml

@@ -67,6 +67,11 @@ system_prompt: |-
      - Note that executed code is not visible to users. If users need to see the code, use '<DISPLAY:language_type>code</DISPLAY>' for displaying code.
      - **IMPORTANT**: After code execution, the system will return content with "Observation:" marker (this is the real execution result). Please continue your next thinking based on these real results. **Do NOT fabricate observation results before code execution.**
 
+  3. Self-verification:
+     - After critical events (tool calls, retrieval results, code execution, agent handoffs, and final-answer preparation), the system may run explicit verification.
+     - If verification reports errors, insufficient evidence, incomplete parameters, or unreliable results, you must repair the issue, gather more evidence, call tools again, or clearly state what cannot be completed.
+     - The final answer is shown to the user only after verification passes. If the system returns Verification feedback, treat it as a real observation and continue revising.
+
   After thinking, when you believe you can answer the user's question, you can generate a final answer directly to the user without generating code and stop the loop.
 
   When generating the final answer, you need to follow these specifications:
@@ -222,3 +227,13 @@ final_answer:
     Original task: {{task}}
 
     Please provide a clear and concise summary of the work completed so far.
+
+
+verification:
+  pre_messages: |-
+    You are a strict verifier for a ReAct agent. Judge reliability only from the task, candidate answer, tool outputs, and observations. Do not output hidden chain-of-thought.
+    You must output JSON only.
+
+  post_messages: |-
+    Verify whether the candidate answer covers the user's intent, is grounded in observations, handles tool errors, uses trustworthy citations, and is formatted for users.
+    Output fields: passed, score, status, failed_criteria, checks, revision_instruction, user_visible_note.

backend/prompts/manager_system_prompt_template_zh.yaml

@@ -130,6 +130,11 @@ system_prompt: |-
      - 注意运行的代码不会被用户看到，所以如果用户需要看到代码，你需要使用'<DISPLAY:语言类型>代码</DISPLAY>'表达展示代码。
      - **重要**：代码执行后，系统会返回 "Observation:" 标记的内容（这是真实的执行结果）。请基于这些真实结果继续下一步思考，**不要在代码执行前自行编造观察结果**。
 
+  3. 自验证：
+     - 关键事件（工具调用、检索结果、代码执行、助手返回、准备最终回答）后，系统会进行显式自验证。
+     - 如果自验证提示存在错误、证据不足、参数不完整或结果不可靠，必须优先修正、补充证据、重新调用工具，或清晰说明无法完成的部分。
+     - 最终回答只有在自验证通过后才会展示给用户；如果系统返回 Verification feedback，请把它视为真实观察结果继续修正，不要忽略。
+
   在思考结束后，当你认为可以回答用户问题，那么可以不生成代码，直接生成最终回答给到用户并停止循环。
 
   生成最终回答时，你需要遵循以下规范：
@@ -299,3 +304,13 @@ final_answer:
     原始任务：{{task}}
 
     请对迄今为止完成的工作进行清晰、简洁的总结。
+
+
+verification:
+  pre_messages: |-
+    你是 ReAct 智能体的严格验证器。请仅根据任务、候选答案、工具输出和观察结果判断答案是否可靠，不要输出隐藏思维链。
+    你必须只输出 JSON。
+
+  post_messages: |-
+    请验证候选答案是否覆盖用户意图、是否有观察结果支撑、是否处理了工具错误、引用是否可信、格式是否适合展示。
+    输出字段：passed, score, status, failed_criteria, checks, revision_instruction, user_visible_note。

backend/services/agent_service.py

@@ -932,6 +932,7 @@ async def update_agent_info_impl(request: AgentInfoRequest, authorization: str =
                 "prompt_template_name": prompt_template_name,
                 "max_steps": request.max_steps,
                 "provide_run_summary": request.provide_run_summary,
+                "verification_config": request.verification_config,
                 "duty_prompt": request.duty_prompt,
                 "constraint_prompt": request.constraint_prompt,
                 "few_shots_prompt": request.few_shots_prompt,
@@ -1314,6 +1315,7 @@ async def export_agent_by_agent_id(agent_id: int, tenant_id: str, user_id: str)
                                           author=agent_info.get("author"),
                                           max_steps=agent_info["max_steps"],
                                           provide_run_summary=agent_info["provide_run_summary"],
+                                          verification_config=agent_info.get("verification_config"),
                                           duty_prompt=agent_info.get(
                                               "duty_prompt"),
                                           constraint_prompt=agent_info.get(
@@ -1468,6 +1470,7 @@ async def import_agent_by_agent_id(
                                          "prompt_template_name": import_agent_info.prompt_template_name or SYSTEM_PROMPT_TEMPLATE_NAME,
                                          "max_steps": import_agent_info.max_steps,
                                          "provide_run_summary": import_agent_info.provide_run_summary,
+                                         "verification_config": getattr(import_agent_info, "verification_config", None),
                                          "duty_prompt": import_agent_info.duty_prompt,
                                          "constraint_prompt": import_agent_info.constraint_prompt,
                                          "few_shots_prompt": import_agent_info.few_shots_prompt,

backend/utils/context_utils.py

@@ -611,6 +611,11 @@ def build_skeleton_execution_flow_component(
         lines.append("   - 注意运行的代码不会被用户看到，所以如果用户需要看到代码，你需要使用'<DISPLAY:语言类型>代码</DISPLAY>'表达展示代码。")
         lines.append("   - **重要**：代码执行后，系统会返回 \"Observation:\" 标记的内容（这是真实的执行结果）。请基于这些真实结果继续下一步思考，**不要在代码执行前自行编造观察结果**。")
         lines.append("")
+        lines.append("3. 自验证：")
+        lines.append("   - 关键事件（工具调用、检索结果、代码执行、助手返回、准备最终回答）后，系统会进行显式自验证。")
+        lines.append("   - 如果自验证提示存在错误、证据不足、参数不完整或结果不可靠，必须优先修正、补充证据、重新调用工具，或清晰说明无法完成的部分。")
+        lines.append("   - 最终回答只有在自验证通过后才会展示给用户；如果系统返回 Verification feedback，请把它视为真实观察结果继续修正，不要忽略。")
+        lines.append("")
         lines.append("在思考结束后，当你认为可以回答用户问题，那么可以不生成代码，直接生成最终回答给到用户并停止循环。")
         lines.append("")
         lines.append("生成最终回答时，你需要遵循以下规范：")
@@ -652,6 +657,11 @@ def build_skeleton_execution_flow_component(
         lines.append("   - Note that executed code is not visible to users. If users need to see the code, use '<DISPLAY:language_type>code</DISPLAY>' for displaying code.")
         lines.append("   - **IMPORTANT**: After code execution, the system will return content with \"Observation:\" marker (this is the real execution result). Please continue your next thinking based on these real results. **Do NOT fabricate observation results before code execution.**")
         lines.append("")
+        lines.append("3. Self-verification:")
+        lines.append("   - After critical events (tool calls, retrieval results, code execution, agent handoffs, and final-answer preparation), the system may run explicit verification.")
+        lines.append("   - If verification reports errors, insufficient evidence, incomplete parameters, or unreliable results, you must repair the issue, gather more evidence, call tools again, or clearly state what cannot be completed.")
+        lines.append("   - The final answer is shown to the user only after verification passes. If the system returns Verification feedback, treat it as a real observation and continue revising.")
+        lines.append("")
         lines.append("After thinking, when you believe you can answer the user's question, you can generate a final answer directly to the user without generating code and stop the loop.")
         lines.append("")
         lines.append("When generating the final answer, you need to follow these specifications:")
@@ -1329,4 +1339,4 @@ def build_app_context_string(
         Formatted app context string
     """
     time_str = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-    return _format_app_context(app_name, app_description, user_id, time_str)
+    return _format_app_context(app_name, app_description, user_id, time_str)

docker/init.sql

@@ -337,6 +337,7 @@ CREATE TABLE IF NOT EXISTS nexent.ag_tenant_agent_t (
     is_new BOOLEAN DEFAULT FALSE,
     provide_run_summary BOOLEAN DEFAULT FALSE,
     enable_context_manager BOOLEAN DEFAULT FALSE,
+    verification_config JSONB,
     version_no INTEGER DEFAULT 0 NOT NULL,
     current_version_no INTEGER NULL,
     ingroup_permission VARCHAR(30),
@@ -397,6 +398,7 @@ COMMENT ON COLUMN nexent.ag_tenant_agent_t.version_no IS 'Version number. 0 = dr
 COMMENT ON COLUMN nexent.ag_tenant_agent_t.current_version_no IS 'Current published version number. NULL means no version published yet';
 COMMENT ON COLUMN nexent.ag_tenant_agent_t.ingroup_permission IS 'In-group permission: EDIT, READ_ONLY, PRIVATE';
 COMMENT ON COLUMN nexent.ag_tenant_agent_t.enable_context_manager IS 'Whether to enable context management (compression) for this agent';
+COMMENT ON COLUMN nexent.ag_tenant_agent_t.verification_config IS 'Layered ReAct self-verification configuration';
 
 -- Create index for is_new queries
 CREATE INDEX IF NOT EXISTS idx_ag_tenant_agent_t_is_new

docker/sql/v2.2.1_0601_add_agent_verification_config.sql

@@ -0,0 +1,7 @@
+-- Migration: Add layered ReAct self-verification config to agents
+-- Description: Stores per-agent verification controls for step-level and final-answer validation.
+
+ALTER TABLE nexent.ag_tenant_agent_t
+ADD COLUMN IF NOT EXISTS verification_config JSONB;
+
+COMMENT ON COLUMN nexent.ag_tenant_agent_t.verification_config IS 'Layered ReAct self-verification configuration';