Skip to content

refactor: opt tool-retry logic #613

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Rasic2 merged 5 commits into from
Dec 31, 2025
Merged

refactor: opt tool-retry logic #613

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f984f89
refactor: remove execution_result_agent and associated prompt for str…
Rasic2 Dec 31, 2025
e140d3b
refactor: remove redundant logger configuration to simplify agent ini…
Rasic2 Dec 31, 2025
b44cee9
refactor: replace StepValidationAgent with DisallowTransferAndContent…
Rasic2 Dec 31, 2025
2a978c3
refactor: streamline code by eliminating unnecessary variables and cl…
Rasic2 Dec 31, 2025
f455732
feat: add tool retry logic with MAX_TOOL_RETRIES for execution agents
Rasic2 Dec 31, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions agents/matmaster_agent/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
USE_PHOTON = False
USER_DIRECT_CONSUME = True
MAX_TOKENS_LIMIT = 128000
MAX_TOOL_RETRIES = 2
13 changes: 1 addition & 12 deletions agents/matmaster_agent/flow_agents/agent.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,9 +31,6 @@
from agents.matmaster_agent.flow_agents.execution_agent.agent import (
MatMasterSupervisorAgent,
)
from agents.matmaster_agent.flow_agents.execution_result_agent.prompt import (
PLAN_EXECUTION_CHECK_INSTRUCTION,
)
from agents.matmaster_agent.flow_agents.expand_agent.agent import ExpandAgent
from agents.matmaster_agent.flow_agents.expand_agent.prompt import EXPAND_INSTRUCTION
from agents.matmaster_agent.flow_agents.expand_agent.schema import ExpandSchema
Expand Down Expand Up @@ -162,13 +159,6 @@ def after_init(self):
# instruction=PLAN_INFO_INSTRUCTION,
)

execution_result_agent = DisallowTransferAndContentLimitLlmAgent(
name='execution_result_agent',
model=MatMasterLlmConfig.gpt_5_mini,
description='汇总计划的执行情况',
instruction=PLAN_EXECUTION_CHECK_INSTRUCTION,
)

self._execution_agent = MatMasterSupervisorAgent(
name='execution_agent',
model=MatMasterLlmConfig.default_litellm_model,
Expand All @@ -177,8 +167,7 @@ def after_init(self):
sub_agents=[
sub_agent(MatMasterLlmConfig)
for sub_agent in AGENT_CLASS_MAPPING.values()
]
+ [execution_result_agent],
],
)

self._analysis_agent = DisallowTransferAndContentLimitLlmAgent(
Expand Down
123 changes: 61 additions & 62 deletions agents/matmaster_agent/flow_agents/execution_agent/agent.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,21 +4,26 @@

from google.adk.agents import InvocationContext
from google.adk.events import Event
from pydantic import model_validator
from opik.integrations.adk import track_adk_agent_recursive
from pydantic import computed_field, model_validator

from agents.matmaster_agent.base_callbacks.public_callback import check_transfer
from agents.matmaster_agent.config import MAX_TOOL_RETRIES
from agents.matmaster_agent.constant import MATMASTER_AGENT_NAME, ModelRole
from agents.matmaster_agent.core_agents.base_agents.schema_agent import (
DisallowTransferAndContentLimitSchemaAgent,
)
from agents.matmaster_agent.core_agents.comp_agents.dntransfer_climit_agent import (
DisallowTransferAndContentLimitLlmAgent,
)
from agents.matmaster_agent.flow_agents.constant import MATMASTER_SUPERVISOR_AGENT
from agents.matmaster_agent.flow_agents.model import PlanStepStatusEnum
from agents.matmaster_agent.flow_agents.step_validation_agent.agent import (
StepValidationAgent,
)
from agents.matmaster_agent.flow_agents.step_validation_agent.prompt import (
STEP_VALIDATION_INSTRUCTION,
)
from agents.matmaster_agent.flow_agents.step_validation_agent.schema import (
StepValidationSchema,
)
from agents.matmaster_agent.flow_agents.style import separate_card
from agents.matmaster_agent.flow_agents.utils import (
check_plan,
Expand Down Expand Up @@ -60,23 +65,37 @@ def after_init(self):
]

# Initialize validation agent
self._validation_agent = StepValidationAgent(state_key='step_validation')
self._validation_agent = DisallowTransferAndContentLimitSchemaAgent(
name='step_validation_agent',
model=MatMasterLlmConfig.tool_schema_model,
description='校验步骤执行结果是否合理',
instruction=STEP_VALIDATION_INSTRUCTION,
output_schema=StepValidationSchema,
state_key='step_validation',
)

self.sub_agents += self.validation_agent
track_adk_agent_recursive(self.validation_agent, MatMasterLlmConfig.opik_tracer)

return self

@computed_field
@property
def validation_agent(self) -> DisallowTransferAndContentLimitSchemaAgent:
return self._validation_agent

@override
async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, None]:
plan = ctx.session.state['plan']
logger.info(f'{ctx.session.id} plan = {plan}')
steps = plan['steps']

for index, step in enumerate(steps):
for index, step in enumerate(plan['steps']):
if step.get('tool_name'):
tried_tools = [step['tool_name']]
current_tool_name = step['tool_name']
tried_tools = [current_tool_name]
alternatives = find_alternative_tool(current_tool_name)
tool_attempt_success = False

tool_attempt_success = False
while not tool_attempt_success:
target_agent = get_agent_name(current_tool_name, self.sub_agents)
logger.info(
Expand All @@ -88,9 +107,10 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
PlanStepStatusEnum.FAILED,
PlanStepStatusEnum.SUBMITTED,
]:
max_retries = 2
retry_count = 0
while retry_count <= max_retries:
validation_reason = ''
# 同一工具重试
while retry_count <= MAX_TOOL_RETRIES:
if step['status'] != PlanStepStatusEnum.SUBMITTED:
update_plan = copy.deepcopy(ctx.session.state['plan'])
update_plan['steps'][index][
Expand Down Expand Up @@ -119,8 +139,8 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
logger.info(
f'{ctx.session.id} Before Run: plan_index = {ctx.session.state["plan_index"]}, plan = {ctx.session.state['plan']}'
)
if retry_count != 0:
separate_card_info = 'ReExecuteSteps'
if retry_count:
separate_card_info = 'ReExecuteStep'
else:
separate_card_info = 'Step'

Expand All @@ -141,36 +161,27 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
)

current_steps = ctx.session.state['plan']['steps']

# 工具调用结果返回【成功】
if (
current_steps[index]['status']
== PlanStepStatusEnum.SUCCESS
):
# 执行结果校验
# 获取执行结果 - 从步骤状态或最近的事件中提取
execution_result = current_steps[index].get(
'result', ''
# 对成功的工具调用结果进行校验
lines = (
f"用户原始请求: {ctx.user_content.parts[0].text}",
f"当前步骤描述: {step['description']}",
f"工具名称: {current_tool_name}",
'请根据以上信息判断，工具的参数配置及对应的执行结果是否严格满足用户原始需求。',
)
if not execution_result:
# 尝试从会话状态获取最后的结果
execution_result = ctx.session.state.get(
'last_execution_result', ''
)

validation_instruction = f"""
用户原始请求: {ctx.user_content.parts[0].text}
当前步骤描述: {step['description']}
工具名称: {current_tool_name}
步骤执行结果: {execution_result}

请根据以上信息判断，工具的参数配置及对应的执行结果是否严格满足用户原始需求。
"""
self._validation_agent.instruction = (
validation_instruction = '\n'.join(lines)
self.validation_agent.instruction = (
STEP_VALIDATION_INSTRUCTION + validation_instruction
)

async for (
validation_event
) in self._validation_agent.run_async(ctx):
) in self.validation_agent.run_async(ctx):
yield validation_event

validation_result = ctx.session.state.get(
Expand All @@ -179,7 +190,8 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
is_valid = validation_result.get('is_valid', True)
validation_reason = validation_result.get('reason', '')

if (not is_valid) and retry_count < max_retries:
# “假成功”结果，计划重试
if (not is_valid) and retry_count < MAX_TOOL_RETRIES:
retry_count += 1
logger.warning(
f'{ctx.session.id} Step {index + 1} validation failed: {validation_reason}'
Expand All @@ -194,7 +206,7 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
):
yield validation_failed_event

# 校验失败，标记为失败状态并准备重试
# 重新标记为进行中状态，准备重试
update_plan = copy.deepcopy(
ctx.session.state['plan']
)
Expand All @@ -211,23 +223,21 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
yield update_state_event(
ctx, state_delta={'plan': update_plan}
)
# 继续循环进行重试
else:
# 校验成功，步骤完成
tool_attempt_success = True
break
elif (
current_steps[index]['status']
== PlanStepStatusEnum.FAILED
and retry_count < max_retries
and retry_count < MAX_TOOL_RETRIES
):
retry_count += 1

# 向用户显示重试信息
retry_message = (
f"步骤 {index + 1} 执行失败，正在准备重试..."
)

for retry_event in all_text_event(
ctx,
self.name,
Expand All @@ -236,44 +246,33 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
):
yield retry_event

update_plan = copy.deepcopy(ctx.session.state['plan'])
update_plan['steps'][index][
'status'
] = PlanStepStatusEnum.PROCESS
if validation_reason:
logger.info(
f'{ctx.session.id} Step {index + 1} failed due to validation, retrying {retry_count}/{max_retries}. Reason: {validation_reason}'
f'{ctx.session.id} Step {index + 1} failed due to validation, retrying {retry_count}/{MAX_TOOL_RETRIES}. Reason: {validation_reason}'
)
# 在重试时更新步骤描述，包含校验失败的原因
update_plan = copy.deepcopy(
ctx.session.state['plan']
)
original_description = step['description']
update_plan['steps'][index][
'description'
] = f"{original_description}\n\n注意：上次执行因以下原因校验失败，请改进：{validation_reason}"
update_plan['steps'][index][
'status'
] = PlanStepStatusEnum.PROCESS
yield update_state_event(
ctx, state_delta={'plan': update_plan}
)
else:
logger.info(
f'{ctx.session.id} Step {index + 1} execution failed, retrying {retry_count}/{max_retries}'
)
# 重置状态为 PROCESS 以便重试
update_plan = copy.deepcopy(
ctx.session.state['plan']
)
update_plan['steps'][index][
'status'
] = PlanStepStatusEnum.PROCESS
yield update_state_event(
ctx, state_delta={'plan': update_plan}
f'{ctx.session.id} Step {index + 1} execution failed, retrying {retry_count}/{MAX_TOOL_RETRIES}'
)

# 重置状态为 PROCESS 以便重试
yield update_state_event(
ctx, state_delta={'plan': update_plan}
)
else:
# 其他状态（SUBMITTED等），退出循环
break

# 如果retry循环结束后仍未成功，尝试alternative工具
# 如果同一工具重试 MAX_TOOL_RETRIES 后仍未成功，尝试其他工具
if not tool_attempt_success:
available_alts = [
alt for alt in alternatives if alt not in tried_tools
Expand All @@ -295,6 +294,7 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
logger.info(
f'{ctx.session.id} Switching to alternative tool: {next_tool} for step {index + 1}'
)

# 更新plan中的tool_name和status
update_plan = copy.deepcopy(ctx.session.state['plan'])
update_plan['steps'][index]['tool_name'] = next_tool
Expand All @@ -313,8 +313,7 @@ async def _run_events(self, ctx: InvocationContext) -> AsyncGenerator[Event, Non
ctx, state_delta={'plan': update_plan}
)
else:

logger.info(
logger.warning(
f'{ctx.session.id} No more alternative tools for step {index + 1}'
)
break # 退出tool while
Expand Down
Empty file removed agents/matmaster_agent/flow_agents/execution_result_agent/__init__.py
View file
Open in desktop
Empty file.
43 changes: 0 additions & 43 deletions agents/matmaster_agent/flow_agents/execution_result_agent/prompt.py
View file
Open in desktop

This file was deleted.

5 changes: 0 additions & 5 deletions agents/matmaster_agent/flow_agents/step_validation_agent/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +0,0 @@
from agents.matmaster_agent.flow_agents.step_validation_agent.agent import (
StepValidationAgent,
)

__all__ = ['StepValidationAgent']
Loading
Loading