Merge branch 'mgx_ops' into 'data_analyst_ldy'

# Conflicts:
#   metagpt/schema.py

metagpt/actions/action.py

@@ -90,7 +90,7 @@ class Action(SerializationMixin, ContextMixin, BaseModel):
         msgs = args[0]
         context = "## History Messages\n"
         context += "\n".join([f"{idx}: {i}" for idx, i in enumerate(reversed(msgs))])
-        return await self.node.fill(context=context, llm=self.llm)
+        return await self.node.fill(req=context, llm=self.llm)
 
     async def run(self, *args, **kwargs):
         """Run action"""

metagpt/actions/action_node.py

@@ -18,6 +18,8 @@ from tenacity import retry, stop_after_attempt, wait_random_exponential
 
 from metagpt.actions.action_outcls_registry import register_action_outcls
 from metagpt.const import MARKDOWN_TITLE_PREFIX, USE_CONFIG_TIMEOUT
+from metagpt.exp_pool import exp_cache
+from metagpt.exp_pool.serializers import ActionNodeSerializer
 from metagpt.llm import BaseLLM
 from metagpt.logs import logger
 from metagpt.provider.postprocess.llm_output_postprocess import llm_output_postprocess
@@ -465,9 +467,11 @@ class ActionNode:
 
         return self
 
+    @exp_cache(serializer=ActionNodeSerializer())
     async def fill(
         self,
-        context,
+        *,
+        req,
         llm,
         schema="json",
         mode="auto",
@@ -478,7 +482,7 @@ class ActionNode:
     ):
         """Fill the node(s) with mode.
 
-        :param context: Everything we should know when filling node.
+        :param req: Everything we should know when filling node.
         :param llm: Large Language Model with pre-defined system message.
         :param schema: json/markdown, determine example and output format.
          - raw: free form text
@@ -497,7 +501,7 @@ class ActionNode:
         :return: self
         """
         self.set_llm(llm)
-        self.set_context(context)
+        self.set_context(req)
         if self.schema:
             schema = self.schema
 

metagpt/actions/design_api.py

@@ -178,12 +178,12 @@ class WriteDesign(Action):
         )
 
     async def _new_system_design(self, context):
-        node = await DESIGN_API_NODE.fill(context=context, llm=self.llm, schema=self.prompt_schema)
+        node = await DESIGN_API_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         return node
 
     async def _merge(self, prd_doc, system_design_doc):
         context = NEW_REQ_TEMPLATE.format(old_design=system_design_doc.content, context=prd_doc.content)
-        node = await REFINED_DESIGN_NODE.fill(context=context, llm=self.llm, schema=self.prompt_schema)
+        node = await REFINED_DESIGN_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         system_design_doc.content = node.instruct_content.model_dump_json()
         return system_design_doc
 

metagpt/actions/generate_questions.py

@@ -22,4 +22,4 @@ class GenerateQuestions(Action):
     name: str = "GenerateQuestions"
 
     async def run(self, context) -> ActionNode:
-        return await QUESTIONS.fill(context=context, llm=self.llm)
+        return await QUESTIONS.fill(req=context, llm=self.llm)

metagpt/actions/prepare_interview.py

@@ -22,4 +22,4 @@ class PrepareInterview(Action):
     name: str = "PrepareInterview"
 
     async def run(self, context):
-        return await QUESTIONS.fill(context=context, llm=self.llm)
+        return await QUESTIONS.fill(req=context, llm=self.llm)

metagpt/actions/project_management.py

@@ -151,12 +151,12 @@ class WriteTasks(Action):
         return task_doc
 
     async def _run_new_tasks(self, context: str):
-        node = await PM_NODE.fill(context, self.llm, schema=self.prompt_schema)
+        node = await PM_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         return node
 
     async def _merge(self, system_design_doc, task_doc) -> Document:
         context = NEW_REQ_TEMPLATE.format(context=system_design_doc.content, old_task=task_doc.content)
-        node = await REFINED_PM_NODE.fill(context, self.llm, schema=self.prompt_schema)
+        node = await REFINED_PM_NODE.fill(req=context, llm=self.llm, schema=self.prompt_schema)
         task_doc.content = node.instruct_content.model_dump_json()
         return task_doc
 

metagpt/actions/write_code_an_draft.py

@@ -578,7 +578,7 @@ class WriteCodeAN(Action):
 
     async def run(self, context):
         self.llm.system_prompt = "You are an outstanding engineer and can implement any code"
-        return await WRITE_MOVE_NODE.fill(context=context, llm=self.llm, schema="json")
+        return await WRITE_MOVE_NODE.fill(req=context, llm=self.llm, schema="json")
 
 
 async def main():

metagpt/actions/write_code_plan_and_change_an.py

@@ -229,7 +229,7 @@ class WriteCodePlanAndChange(Action):
             code=await self.get_old_codes(),
         )
         logger.info("Writing code plan and change..")
-        return await WRITE_CODE_PLAN_AND_CHANGE_NODE.fill(context=context, llm=self.llm, schema="json")
+        return await WRITE_CODE_PLAN_AND_CHANGE_NODE.fill(req=context, llm=self.llm, schema="json")
 
     async def get_old_codes(self) -> str:
         old_codes = await self.repo.srcs.get_all()

metagpt/actions/write_prd.py

@@ -211,7 +211,7 @@ class WritePRD(Action):
         context = CONTEXT_TEMPLATE.format(requirements=requirement, project_name=project_name)
         exclude = [PROJECT_NAME.key] if project_name else []
         node = await WRITE_PRD_NODE.fill(
-            context=context, llm=self.llm, exclude=exclude, schema=self.prompt_schema
+            req=context, llm=self.llm, exclude=exclude, schema=self.prompt_schema
         )  # schema=schema
         return node
 
@@ -238,7 +238,7 @@ class WritePRD(Action):
     async def _is_bugfix(self, context: str) -> bool:
         if not self.repo.code_files_exists():
             return False
-        node = await WP_ISSUE_TYPE_NODE.fill(context, self.llm)
+        node = await WP_ISSUE_TYPE_NODE.fill(req=context, llm=self.llm)
         return node.get("issue_type") == "BUG"
 
     async def get_related_docs(self, req: Document, docs: list[Document]) -> list[Document]:
@@ -248,14 +248,14 @@ class WritePRD(Action):
 
     async def _is_related(self, req: Document, old_prd: Document) -> bool:
         context = NEW_REQ_TEMPLATE.format(old_prd=old_prd.content, requirements=req.content)
-        node = await WP_IS_RELATIVE_NODE.fill(context, self.llm)
+        node = await WP_IS_RELATIVE_NODE.fill(req=context, llm=self.llm)
         return node.get("is_relative") == "YES"
 
     async def _merge(self, req: Document, related_doc: Document) -> Document:
         if not self.project_name:
             self.project_name = Path(self.project_path).name
         prompt = NEW_REQ_TEMPLATE.format(requirements=req.content, old_prd=related_doc.content)
-        node = await REFINED_PRD_NODE.fill(context=prompt, llm=self.llm, schema=self.prompt_schema)
+        node = await REFINED_PRD_NODE.fill(req=prompt, llm=self.llm, schema=self.prompt_schema)
         related_doc.content = node.instruct_content.model_dump_json()
         await self._rename_workspace(node)
         return related_doc

metagpt/actions/write_review.py

@@ -36,4 +36,4 @@ class WriteReview(Action):
     name: str = "WriteReview"
 
     async def run(self, context):
-        return await WRITE_REVIEW_NODE.fill(context=context, llm=self.llm, schema="json")
+        return await WRITE_REVIEW_NODE.fill(req=context, llm=self.llm, schema="json")

metagpt/config2.py

@@ -13,6 +13,7 @@ from pydantic import BaseModel, model_validator
 
 from metagpt.configs.browser_config import BrowserConfig
 from metagpt.configs.embedding_config import EmbeddingConfig
+from metagpt.configs.exp_pool_config import ExperiencePoolConfig
 from metagpt.configs.llm_config import LLMConfig, LLMType
 from metagpt.configs.mermaid_config import MermaidConfig
 from metagpt.configs.redis_config import RedisConfig
@@ -71,6 +72,9 @@ class Config(CLIParams, YamlModel):
     enable_longterm_memory: bool = False
     code_review_k_times: int = 2
 
+    # Experience Pool Parameters
+    exp_pool: ExperiencePoolConfig = ExperiencePoolConfig()
+
     # Will be removed in the future
     metagpt_tti_url: str = ""
     language: str = "English"

metagpt/configs/exp_pool_config.py

@@ -0,0 +1,9 @@
+from pydantic import Field
+
+from metagpt.utils.yaml_model import YamlModel
+
+
+class ExperiencePoolConfig(YamlModel):
+    enable_read: bool = Field(default=False, description="Enable to read from experience pool.")
+    enable_write: bool = Field(default=False, description="Enable to write to experience pool.")
+    persist_path: str = Field(default=".chroma_exp_data", description="The persist path for experience pool.")

metagpt/environment/mgx/mgx_env.py

@@ -27,7 +27,7 @@ class MGXEnv(Environment):
 
     def publish_message(self, message: Message, user_defined_recipient: str = "", publicer: str = "") -> bool:
         """let the team leader take over message publishing"""
-        tl = self.get_role("Tim")  # TeamLeader's name is Tim
+        tl = self.get_role("Mike")  # TeamLeader's name is Mike
 
         if user_defined_recipient:
             # human user's direct chat message to a certain role

metagpt/exp_pool/__init__.py

@@ -0,0 +1,6 @@
+"""Experience pool init."""
+
+from metagpt.exp_pool.manager import exp_manager
+from metagpt.exp_pool.decorator import exp_cache
+
+__all__ = ["exp_manager", "exp_cache"]

metagpt/exp_pool/context_builders/__init__.py

@@ -0,0 +1,7 @@
+"""Context builders init."""
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+from metagpt.exp_pool.context_builders.simple import SimpleContextBuilder
+from metagpt.exp_pool.context_builders.role_zero import RoleZeroContextBuilder
+
+__all__ = ["BaseContextBuilder", "SimpleContextBuilder", "RoleZeroContextBuilder"]

metagpt/exp_pool/context_builders/action_node.py

@@ -0,0 +1,30 @@
+"""Action Node context builder."""
+
+from typing import Any
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+
+ACTION_NODE_CONTEXT_TEMPLATE = """
+{req}
+
+### Experiences
+-----
+{exps}
+-----
+
+## Instruction
+Consider **Experiences** to generate a better answer.
+"""
+
+
+class ActionNodeContextBuilder(BaseContextBuilder):
+    async def build(self, req: Any) -> str:
+        """Builds the action node context string.
+
+        If there are no experiences, returns the original `req`;
+        otherwise returns context with `req` and formatted experiences.
+        """
+
+        exps = self.format_exps()
+
+        return ACTION_NODE_CONTEXT_TEMPLATE.format(req=req, exps=exps) if exps else req

metagpt/exp_pool/context_builders/base.py

@@ -0,0 +1,41 @@
+"""Base context builder."""
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Experience
+
+EXP_TEMPLATE = """Given the request: {req}, We can get the response: {resp}, which scored: {score}."""
+
+
+class BaseContextBuilder(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    exps: list[Experience] = []
+
+    @abstractmethod
+    async def build(self, req: Any) -> Any:
+        """Build context from req.
+
+        Do not modify `req`. If modification is necessary, use copy.deepcopy to create a copy first.
+        """
+
+    def format_exps(self) -> str:
+        """Format experiences into a numbered list of strings.
+
+        Example:
+            1. Given the request: req1, We can get the response: resp1, which scored: 8.
+            2. Given the request: req2, We can get the response: resp2, which scored: 9.
+
+        Returns:
+            str: The formatted experiences as a string.
+        """
+
+        result = []
+        for i, exp in enumerate(self.exps, start=1):
+            score_val = exp.metric.score.val if exp.metric and exp.metric.score else "N/A"
+            result.append(f"{i}. " + EXP_TEMPLATE.format(req=exp.req, resp=exp.resp, score=score_val))
+
+        return "\n".join(result)

metagpt/exp_pool/context_builders/role_zero.py

@@ -0,0 +1,56 @@
+"""RoleZero context builder."""
+
+import copy
+import re
+from typing import Any
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+
+
+class RoleZeroContextBuilder(BaseContextBuilder):
+    async def build(self, req: Any) -> list[dict]:
+        """Builds the role zero context string.
+
+        Note:
+            1. The expected format for `req`, e.g., [{...}, {"role": "user", "content": "context"}].
+            2. Returns the original `req` if it is empty.
+            3. Creates a copy of req and replaces the example content in the copied req with actual experiences.
+        """
+
+        if not req:
+            return req
+
+        exps = self.format_exps()
+        if not exps:
+            return req
+
+        req_copy = copy.deepcopy(req)
+
+        req_copy[-1]["content"] = self.replace_example_content(req_copy[-1].get("content", ""), exps)
+
+        return req_copy
+
+    def replace_example_content(self, text: str, new_example_content: str) -> str:
+        return self.replace_content_between_markers(text, "# Example", "# Instruction", new_example_content)
+
+    @staticmethod
+    def replace_content_between_markers(text: str, start_marker: str, end_marker: str, new_content: str) -> str:
+        """Replace the content between `start_marker` and `end_marker` in the text with `new_content`.
+
+        Args:
+            text (str): The original text.
+            new_content (str): The new content to replace the old content.
+            start_marker (str): The marker indicating the start of the content to be replaced, such as '# Example'.
+            end_marker (str): The marker indicating the end of the content to be replaced, such as '# Instruction'.
+
+        Returns:
+            str: The text with the content replaced.
+        """
+
+        pattern = re.compile(f"({start_marker}\n)(.*?)(\n{end_marker})", re.DOTALL)
+
+        def replacement(match):
+            return f"{match.group(1)}{new_content}\n{match.group(3)}"
+
+        replaced_text = pattern.sub(replacement, text)
+        return replaced_text

metagpt/exp_pool/context_builders/simple.py

@@ -0,0 +1,26 @@
+"""Simple context builder."""
+
+
+from typing import Any
+
+from metagpt.exp_pool.context_builders.base import BaseContextBuilder
+
+SIMPLE_CONTEXT_TEMPLATE = """
+## Context
+
+### Experiences
+-----
+{exps}
+-----
+
+## User Requirement
+{req}
+
+## Instruction
+Consider **Experiences** to generate a better answer.
+"""
+
+
+class SimpleContextBuilder(BaseContextBuilder):
+    async def build(self, req: Any) -> str:
+        return SIMPLE_CONTEXT_TEMPLATE.format(req=req, exps=self.format_exps())

metagpt/exp_pool/decorator.py

@@ -0,0 +1,211 @@
+"""Experience Decorator."""
+
+import asyncio
+import functools
+from typing import Any, Callable, Optional, TypeVar
+
+from pydantic import BaseModel, ConfigDict, model_validator
+
+from metagpt.config2 import config
+from metagpt.exp_pool.context_builders import BaseContextBuilder, SimpleContextBuilder
+from metagpt.exp_pool.manager import ExperienceManager, exp_manager
+from metagpt.exp_pool.perfect_judges import BasePerfectJudge, SimplePerfectJudge
+from metagpt.exp_pool.schema import Experience, Metric, QueryType, Score
+from metagpt.exp_pool.scorers import BaseScorer, SimpleScorer
+from metagpt.exp_pool.serializers import BaseSerializer, SimpleSerializer
+from metagpt.logs import logger
+from metagpt.utils.async_helper import NestAsyncio
+from metagpt.utils.exceptions import handle_exception
+
+ReturnType = TypeVar("ReturnType")
+
+
+def exp_cache(
+    _func: Optional[Callable[..., ReturnType]] = None,
+    query_type: QueryType = QueryType.SEMANTIC,
+    manager: Optional[ExperienceManager] = None,
+    scorer: Optional[BaseScorer] = None,
+    perfect_judge: Optional[BasePerfectJudge] = None,
+    context_builder: Optional[BaseContextBuilder] = None,
+    serializer: Optional[BaseSerializer] = None,
+    tag: Optional[str] = None,
+):
+    """Decorator to get a perfect experience, otherwise, it executes the function, and create a new experience.
+
+    Note:
+        1. This can be applied to both synchronous and asynchronous functions.
+        2. The function must have a `req` parameter, and it must be provided as a keyword argument.
+        3. If `config.exp_pool.enable_read` is False, the decorator will just directly execute the function.
+        4. If `config.exp_pool.enable_write` is False, the decorator will skip evaluating and saving the experience.
+
+    Args:
+        _func: Just to make the decorator more flexible, for example, it can be used directly with @exp_cache by default, without the need for @exp_cache().
+        query_type: The type of query to be used when fetching experiences.
+        manager: How to fetch, evaluate and save experience, etc. Default to `exp_manager`.
+        scorer: Evaluate experience. Default to `SimpleScorer()`.
+        perfect_judge: Determines if an experience is perfect. Defaults to `SimplePerfectJudge()`.
+        context_builder: Build the context from exps and the function parameters. Default to `SimpleContextBuilder()`.
+        serializer: Serializes the request and the function's return value for storage, deserializes the stored response back to the function's return value. Defaults to `SimpleSerializer()`.
+        tag: An optional tag for the experience. Default to `ClassName.method_name` or `function_name`.
+    """
+
+    def decorator(func: Callable[..., ReturnType]) -> Callable[..., ReturnType]:
+        if not config.exp_pool.enable_read:
+            return func
+
+        @functools.wraps(func)
+        async def get_or_create(args: Any, kwargs: Any) -> ReturnType:
+            handler = ExpCacheHandler(
+                func=func,
+                args=args,
+                kwargs=kwargs,
+                query_type=query_type,
+                exp_manager=manager,
+                exp_scorer=scorer,
+                exp_perfect_judge=perfect_judge,
+                context_builder=context_builder,
+                serializer=serializer,
+                tag=tag,
+            )
+
+            await handler.fetch_experiences()
+
+            if exp := await handler.get_one_perfect_exp():
+                return exp
+
+            await handler.execute_function()
+
+            if config.exp_pool.enable_write:
+                await handler.process_experience()
+
+            return handler._raw_resp
+
+        return ExpCacheHandler.choose_wrapper(func, get_or_create)
+
+    return decorator(_func) if _func else decorator
+
+
+class ExpCacheHandler(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    func: Callable
+    args: Any
+    kwargs: Any
+    query_type: QueryType = QueryType.SEMANTIC
+    exp_manager: Optional[ExperienceManager] = None
+    exp_scorer: Optional[BaseScorer] = None
+    exp_perfect_judge: Optional[BasePerfectJudge] = None
+    context_builder: Optional[BaseContextBuilder] = None
+    serializer: Optional[BaseSerializer] = None
+    tag: Optional[str] = None
+
+    _exps: list[Experience] = None
+    _req: str = ""
+    _resp: str = ""
+    _raw_resp: Any = None
+    _score: Score = None
+
+    @model_validator(mode="after")
+    def initialize(self):
+        """Initialize default values for optional parameters if they are None.
+
+        This is necessary because the decorator might pass None, which would override the default values set by Field.
+        """
+
+        self._validate_params()
+
+        self.exp_manager = self.exp_manager or exp_manager
+        self.exp_scorer = self.exp_scorer or SimpleScorer()
+        self.exp_perfect_judge = self.exp_perfect_judge or SimplePerfectJudge()
+        self.context_builder = self.context_builder or SimpleContextBuilder()
+        self.serializer = self.serializer or SimpleSerializer()
+        self.tag = self.tag or self._generate_tag()
+
+        self._req = self.serializer.serialize_req(**self.kwargs)
+
+        return self
+
+    async def fetch_experiences(self):
+        """Fetch experiences by query_type."""
+
+        self._exps = await self.exp_manager.query_exps(self._req, query_type=self.query_type, tag=self.tag)
+
+    async def get_one_perfect_exp(self) -> Optional[Any]:
+        """Get a potentially perfect experience, and resolve resp."""
+
+        for exp in self._exps:
+            if await self.exp_perfect_judge.is_perfect_exp(exp, self._req, *self.args, **self.kwargs):
+                logger.info(f"Get one perfect experience: {exp.req[:20]}...")
+                return self.serializer.deserialize_resp(exp.resp)
+
+        return None
+
+    async def execute_function(self):
+        """Execute the function, and save resp."""
+
+        self._raw_resp = await self._execute_function()
+        self._resp = self.serializer.serialize_resp(self._raw_resp)
+
+    @handle_exception
+    async def process_experience(self):
+        """Process experience.
+
+        Evaluates and saves experience.
+        Use `handle_exception` to ensure robustness, do not stop subsequent operations.
+        """
+
+        await self.evaluate_experience()
+        self.save_experience()
+
+    async def evaluate_experience(self):
+        """Evaluate the experience, and save the score."""
+
+        self._score = await self.exp_scorer.evaluate(self._req, self._resp)
+
+    def save_experience(self):
+        """Save the new experience."""
+
+        exp = Experience(req=self._req, resp=self._resp, tag=self.tag, metric=Metric(score=self._score))
+        self.exp_manager.create_exp(exp)
+
+    @staticmethod
+    def choose_wrapper(func, wrapped_func):
+        """Choose how to run wrapped_func based on whether the function is asynchronous."""
+
+        async def async_wrapper(*args, **kwargs):
+            return await wrapped_func(args, kwargs)
+
+        def sync_wrapper(*args, **kwargs):
+            NestAsyncio.apply_once()
+            return asyncio.get_event_loop().run_until_complete(wrapped_func(args, kwargs))
+
+        return async_wrapper if asyncio.iscoroutinefunction(func) else sync_wrapper
+
+    def _validate_params(self):
+        if "req" not in self.kwargs:
+            raise ValueError("`req` must be provided as a keyword argument.")
+
+    def _generate_tag(self) -> str:
+        """Generates a tag for the self.func.
+
+        "ClassName.method_name" if the first argument is a class instance, otherwise just "function_name".
+        """
+
+        if self.args and hasattr(self.args[0], "__class__"):
+            cls_name = type(self.args[0]).__name__
+            return f"{cls_name}.{self.func.__name__}"
+
+        return self.func.__name__
+
+    async def _build_context(self) -> str:
+        self.context_builder.exps = self._exps
+
+        return await self.context_builder.build(self.kwargs["req"])
+
+    async def _execute_function(self):
+        self.kwargs["req"] = await self._build_context()
+
+        if asyncio.iscoroutinefunction(self.func):
+            return await self.func(*self.args, **self.kwargs)
+
+        return self.func(*self.args, **self.kwargs)

metagpt/exp_pool/manager.py

@@ -0,0 +1,116 @@
+"""Experience Manager."""
+
+from typing import TYPE_CHECKING, Any
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.config2 import Config, config
+from metagpt.exp_pool.schema import (
+    DEFAULT_COLLECTION_NAME,
+    DEFAULT_SIMILARITY_TOP_K,
+    Experience,
+    QueryType,
+)
+from metagpt.logs import logger
+from metagpt.utils.exceptions import handle_exception
+
+if TYPE_CHECKING:
+    from llama_index.vector_stores.chroma import ChromaVectorStore
+
+
+class ExperienceManager(BaseModel):
+    """ExperienceManager manages the lifecycle of experiences, including CRUD and optimization.
+
+    Args:
+        config (Config): Configuration for managing experiences.
+        _storage (SimpleEngine): Engine to handle the storage and retrieval of experiences.
+        _vector_store (ChromaVectorStore): The actual place where vectors are stored.
+    """
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    config: Config = config
+
+    _storage: Any = None
+    _vector_store: Any = None
+
+    @property
+    def storage(self):
+        if self._storage is None:
+            try:
+                from metagpt.rag.engines import SimpleEngine
+                from metagpt.rag.schema import ChromaRetrieverConfig, LLMRankerConfig
+            except ImportError:
+                raise ImportError("To use the experience pool, you need to install the rag module.")
+
+            retriever_configs = [
+                ChromaRetrieverConfig(
+                    persist_path=self.config.exp_pool.persist_path,
+                    collection_name=DEFAULT_COLLECTION_NAME,
+                    similarity_top_k=DEFAULT_SIMILARITY_TOP_K,
+                )
+            ]
+            ranker_configs = [LLMRankerConfig(top_n=DEFAULT_SIMILARITY_TOP_K)]
+
+            self._storage: SimpleEngine = SimpleEngine.from_objs(
+                retriever_configs=retriever_configs, ranker_configs=ranker_configs
+            )
+            logger.debug(f"exp_pool config: {self.config.exp_pool}")
+
+        return self._storage
+
+    @property
+    def vector_store(self):
+        if not self._vector_store:
+            self._vector_store: ChromaVectorStore = self.storage._retriever._vector_store
+
+        return self._vector_store
+
+    @handle_exception
+    def create_exp(self, exp: Experience):
+        """Adds an experience to the storage if writing is enabled.
+
+        Args:
+            exp (Experience): The experience to add.
+        """
+
+        if not self.config.exp_pool.enable_write:
+            return
+
+        self.storage.add_objs([exp])
+
+    @handle_exception(default_return=[])
+    async def query_exps(self, req: str, tag: str = "", query_type: QueryType = QueryType.SEMANTIC) -> list[Experience]:
+        """Retrieves and filters experiences.
+
+        Args:
+            req (str): The query string to retrieve experiences.
+            tag (str): Optional tag to filter the experiences by.
+            query_type (QueryType): Default semantic to vector matching. exact to same matching.
+
+        Returns:
+            list[Experience]: A list of experiences that match the args.
+        """
+
+        if not self.config.exp_pool.enable_read:
+            return []
+
+        nodes = await self.storage.aretrieve(req)
+        exps: list[Experience] = [node.metadata["obj"] for node in nodes]
+
+        # TODO: filter by metadata
+        if tag:
+            exps = [exp for exp in exps if exp.tag == tag]
+
+        if query_type == QueryType.EXACT:
+            exps = [exp for exp in exps if exp.req == req]
+
+        return exps
+
+    def get_exps_count(self) -> int:
+        """Get the total number of experiences."""
+
+        return self.vector_store._collection.count()
+
+
+exp_manager = ExperienceManager()

metagpt/exp_pool/perfect_judges/__init__.py

@@ -0,0 +1,6 @@
+"""Perfect judges init."""
+
+from metagpt.exp_pool.perfect_judges.base import BasePerfectJudge
+from metagpt.exp_pool.perfect_judges.simple import SimplePerfectJudge
+
+__all__ = ["BasePerfectJudge", "SimplePerfectJudge"]

metagpt/exp_pool/perfect_judges/base.py

@@ -0,0 +1,20 @@
+"""Base perfect judge."""
+
+from abc import ABC, abstractmethod
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Experience
+
+
+class BasePerfectJudge(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    async def is_perfect_exp(self, exp: Experience, serialized_req: str, *args, **kwargs) -> bool:
+        """Determine whether the experience is perfect.
+
+        Args:
+            exp (Experience): The experience to evaluate.
+            serialized_req (str): The serialized request to compare against the experience's request.
+        """

metagpt/exp_pool/perfect_judges/simple.py

@@ -0,0 +1,27 @@
+"""Simple perfect judge."""
+
+
+from pydantic import ConfigDict
+
+from metagpt.exp_pool.perfect_judges.base import BasePerfectJudge
+from metagpt.exp_pool.schema import MAX_SCORE, Experience
+
+
+class SimplePerfectJudge(BasePerfectJudge):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    async def is_perfect_exp(self, exp: Experience, serialized_req: str, *args, **kwargs) -> bool:
+        """Determine whether the experience is perfect.
+
+        Args:
+            exp (Experience): The experience to evaluate.
+            serialized_req (str): The serialized request to compare against the experience's request.
+
+        Returns:
+            bool: True if the serialized request matches the experience's request and the experience's score is perfect, False otherwise.
+        """
+
+        if not exp.metric or not exp.metric.score:
+            return False
+
+        return serialized_req == exp.req and exp.metric.score.val == MAX_SCORE

metagpt/exp_pool/schema.py

@@ -0,0 +1,72 @@
+"""Experience schema."""
+
+from enum import Enum
+from typing import Optional
+
+from pydantic import BaseModel, Field
+
+MAX_SCORE = 10
+
+DEFAULT_COLLECTION_NAME = "experience_pool"
+DEFAULT_SIMILARITY_TOP_K = 2
+
+
+class QueryType(str, Enum):
+    """Type of query experiences."""
+
+    EXACT = "exact"
+    SEMANTIC = "semantic"
+
+
+class ExperienceType(str, Enum):
+    """Experience Type."""
+
+    SUCCESS = "success"
+    FAILURE = "failure"
+    INSIGHT = "insight"
+
+
+class EntryType(Enum):
+    """Experience Entry Type."""
+
+    AUTOMATIC = "Automatic"
+    MANUAL = "Manual"
+
+
+class Score(BaseModel):
+    """Score in Metric."""
+
+    val: int = Field(default=1, description="Value of the score, Between 1 and 10, higher is better.")
+    reason: str = Field(default="", description="Reason for the value.")
+
+
+class Metric(BaseModel):
+    """Experience Metric."""
+
+    time_cost: float = Field(default=0.000, description="Time cost, the unit is milliseconds.")
+    money_cost: float = Field(default=0.000, description="Money cost, the unit is US dollars.")
+    score: Score = Field(default=None, description="Score, with value and reason.")
+
+
+class Trajectory(BaseModel):
+    """Experience Trajectory."""
+
+    plan: str = Field(default="", description="The plan.")
+    action: str = Field(default="", description="Action for the plan.")
+    observation: str = Field(default="", description="Output of the action.")
+    reward: int = Field(default=0, description="Measure the action.")
+
+
+class Experience(BaseModel):
+    """Experience."""
+
+    req: str = Field(..., description="")
+    resp: str = Field(..., description="The type is string/json/code.")
+    metric: Optional[Metric] = Field(default=None, description="Metric.")
+    exp_type: ExperienceType = Field(default=ExperienceType.SUCCESS, description="The type of experience.")
+    entry_type: EntryType = Field(default=EntryType.AUTOMATIC, description="Type of entry: Manual or Automatic.")
+    tag: str = Field(default="", description="Tagging experience.")
+    traj: Optional[Trajectory] = Field(default=None, description="Trajectory.")
+
+    def rag_key(self):
+        return self.req

metagpt/exp_pool/scorers/__init__.py

@@ -0,0 +1,6 @@
+"""Scorers init."""
+
+from metagpt.exp_pool.scorers.base import BaseScorer
+from metagpt.exp_pool.scorers.simple import SimpleScorer
+
+__all__ = ["BaseScorer", "SimpleScorer"]

metagpt/exp_pool/scorers/base.py

@@ -0,0 +1,15 @@
+"""Base scorer."""
+
+from abc import ABC, abstractmethod
+
+from pydantic import BaseModel, ConfigDict
+
+from metagpt.exp_pool.schema import Score
+
+
+class BaseScorer(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    async def evaluate(self, req: str, resp: str) -> Score:
+        """Evaluates the quality of a response relative to a given request."""

metagpt/exp_pool/scorers/simple.py

@@ -0,0 +1,65 @@
+"""Simple scorer."""
+
+import json
+
+from pydantic import Field
+
+from metagpt.exp_pool.schema import Score
+from metagpt.exp_pool.scorers.base import BaseScorer
+from metagpt.llm import LLM
+from metagpt.provider.base_llm import BaseLLM
+from metagpt.utils.common import CodeParser
+
+SIMPLE_SCORER_TEMPLATE = """
+Role: You are a highly efficient assistant, tasked with evaluating a response to a given request. The response is generated by a large language model (LLM). 
+
+I will provide you with a request and a corresponding response. Your task is to assess this response and provide a score from a human perspective.
+
+## Context
+### Request
+{req}
+
+### Response
+{resp}
+
+## Format Example
+```json
+{{
+    "val": "the value of the score, int from 1 to 10, higher is better.",
+    "reason": "an explanation supporting the score."
+}}
+```
+
+## Instructions
+- Understand the request and response given by the user.
+- Evaluate the response based on its quality relative to the given request.
+- Provide a score from 1 to 10, where 10 is the best.
+- Provide a reason supporting your score.
+
+## Constraint
+Format: Just print the result in json format like **Format Example**.
+
+## Action
+Follow instructions, generate output and make sure it follows the **Constraint**.
+"""
+
+
+class SimpleScorer(BaseScorer):
+    llm: BaseLLM = Field(default_factory=LLM)
+
+    async def evaluate(self, req: str, resp: str) -> Score:
+        """Evaluates the quality of a response relative to a given request, as scored by an LLM.
+
+        Args:
+            req (str): The request.
+            resp (str): The response.
+
+        Returns:
+            Score: An object containing the score (1-10) and the reasoning.
+        """
+
+        prompt = SIMPLE_SCORER_TEMPLATE.format(req=req, resp=resp)
+        resp = await self.llm.aask(prompt)
+        resp_json = json.loads(CodeParser.parse_code(resp, lang="json"))
+
+        return Score(**resp_json)

metagpt/exp_pool/serializers/__init__.py

@@ -0,0 +1,9 @@
+"""Serializers init."""
+
+from metagpt.exp_pool.serializers.base import BaseSerializer
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+from metagpt.exp_pool.serializers.action_node import ActionNodeSerializer
+from metagpt.exp_pool.serializers.role_zero import RoleZeroSerializer
+
+
+__all__ = ["BaseSerializer", "SimpleSerializer", "ActionNodeSerializer", "RoleZeroSerializer"]

metagpt/exp_pool/serializers/action_node.py

@@ -0,0 +1,36 @@
+"""ActionNode Serializer."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Type
+
+# Import ActionNode only for type checking to avoid circular imports
+if TYPE_CHECKING:
+    from metagpt.actions.action_node import ActionNode
+
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+
+
+class ActionNodeSerializer(SimpleSerializer):
+    def serialize_resp(self, resp: ActionNode) -> str:
+        return resp.instruct_content.model_dump_json()
+
+    def deserialize_resp(self, resp: str) -> ActionNode:
+        """Customized deserialization, it will be triggered when a perfect experience is found.
+
+        ActionNode cannot be serialized, it throws an error 'cannot pickle 'SSLContext' object'.
+        """
+
+        class InstructContent:
+            def __init__(self, json_data):
+                self.json_data = json_data
+
+            def model_dump_json(self):
+                return self.json_data
+
+        from metagpt.actions.action_node import ActionNode
+
+        action_node = ActionNode(key="", expected_type=Type[str], instruction="", example="")
+        action_node.instruct_content = InstructContent(resp)
+
+        return action_node

metagpt/exp_pool/serializers/base.py

@@ -0,0 +1,29 @@
+"""Base serializer."""
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict
+
+
+class BaseSerializer(BaseModel, ABC):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+    @abstractmethod
+    def serialize_req(self, **kwargs) -> str:
+        """Serializes the request for storage.
+
+        Do not modify kwargs. If modification is necessary, use copy.deepcopy to create a copy first.
+        Note that copy.deepcopy may raise errors, such as TypeError: cannot pickle '_thread.RLock' object.
+        """
+
+    @abstractmethod
+    def serialize_resp(self, resp: Any) -> str:
+        """Serializes the function's return value for storage.
+
+        Do not modify resp. The rest is the same as `serialize_req`.
+        """
+
+    @abstractmethod
+    def deserialize_resp(self, resp: str) -> Any:
+        """Deserializes the stored response back to the function's return value"""

metagpt/exp_pool/serializers/role_zero.py

@@ -0,0 +1,58 @@
+"""RoleZero Serializer."""
+
+import copy
+import json
+
+from metagpt.exp_pool.serializers.simple import SimpleSerializer
+
+
+class RoleZeroSerializer(SimpleSerializer):
+    def serialize_req(self, **kwargs) -> str:
+        """Serialize the request for database storage, ensuring it is a string.
+
+        Only extracts the necessary content from `req` because `req` may be very lengthy and could cause embedding errors.
+
+        Args:
+            req (list[dict]): The request to be serialized. Example:
+                [
+                    {"role": "user", "content": "..."},
+                    {"role": "assistant", "content": "..."},
+                    {"role": "user", "content": "context"},
+                ]
+
+        Returns:
+            str: The serialized request as a JSON string.
+        """
+        req = kwargs.get("req", [])
+
+        if not req:
+            return ""
+
+        filtered_req = self._filter_req(req)
+
+        if state_data := kwargs.get("state_data"):
+            filtered_req.append({"role": "user", "content": state_data})
+
+        return json.dumps(filtered_req)
+
+    def _filter_req(self, req: list[dict]) -> list[dict]:
+        """Filter the `req` to include only necessary items.
+
+        Args:
+            req (list[dict]): The original request.
+
+        Returns:
+            list[dict]: The filtered request.
+        """
+
+        filtered_req = [copy.deepcopy(item) for item in req if self._is_useful_content(item["content"])]
+
+        return filtered_req
+
+    def _is_useful_content(self, content: str) -> bool:
+        """Currently, only the content of the file is considered, and more judgments can be added later."""
+
+        if "Command Editor.read executed: file_path" in content:
+            return True
+
+        return False

metagpt/exp_pool/serializers/simple.py

@@ -0,0 +1,22 @@
+"""Simple Serializer."""
+
+from typing import Any
+
+from metagpt.exp_pool.serializers.base import BaseSerializer
+
+
+class SimpleSerializer(BaseSerializer):
+    def serialize_req(self, **kwargs) -> str:
+        """Just use `str` to convert the request object into a string."""
+
+        return str(kwargs.get("req", ""))
+
+    def serialize_resp(self, resp: Any) -> str:
+        """Just use `str` to convert the response object into a string."""
+
+        return str(resp)
+
+    def deserialize_resp(self, resp: str) -> Any:
+        """Just return the string response as it is."""
+
+        return resp

metagpt/ext/cr/actions/code_review.py

@@ -5,7 +5,7 @@
 import json
 import re
 
-from unidiff import PatchedFile, PatchSet
+from unidiff import PatchSet
 
 from metagpt.actions.action import Action
 from metagpt.ext.cr.utils.cleaner import (
@@ -19,6 +19,7 @@ from metagpt.utils.common import parse_json_code_block
 
 CODE_REVIEW_PROMPT_TEMPLATE = """
 NOTICE
+Let's think and work step by step.
 With the given pull-request(PR) Patch, and referenced Points(Code Standards), you should compare each point with the code one-by-one.
 
 The Patch code has added line number at the first character each line for reading, but the review should focus on new added code inside the `Patch` (lines starting with line number and '+').
@@ -52,12 +53,13 @@ CodeReview guidelines:
 - Don't suggest to add docstring unless it's necessary indeed.
 - If the same code error occurs multiple times, it cannot be omitted, and all places need to be identified.But Don't duplicate at the same place with the same comment!
 - Every line of code in the patch needs to be carefully checked, and laziness cannot be omitted. It is necessary to find out all the places.
+- The `comment` and `point_id` in the Output must correspond to and belong to the same one `Point`.
 
 Just print the PR Patch comments in json format like **Output Format**.
 """
 
 CODE_REVIEW_COMFIRM_SYSTEM_PROMPT = """
-You are a professional engineer with Java stack, and good at code review comment result judgement.
+You are a professional engineer with {code_language} stack, and good at code review comment result judgement.Let's think and work step by step.
 """
 
 CODE_REVIEW_COMFIRM_TEMPLATE = """
@@ -76,7 +78,7 @@ CODE_REVIEW_COMFIRM_TEMPLATE = """
 
 ## Your Task:
 1. First, check if the code meets the requirements and does not violate any defects. If it meets the requirements and does not violate any defects, print `False` and do not proceed with further judgment.
-2. If the check in step 1 does not print `False`, proceed to further judgment. Based on the "Reference Example for Judgment" provided, determine if the "Code" and "Code Review Comments" match. If they match, print `True`; otherwise, print `False`.
+2. Based on the `Reference Example for Judgment` provided, determine if the `Code` and `Code Review Comments` match. If they match, print `True`; otherwise, print `False`.
 
 Note: Your output should only be `True` or `False` without any explanations.
 """
@@ -89,25 +91,35 @@ class CodeReview(Action):
         new_comments = []
         logger.debug(f"original comments: {comments}")
         for cmt in comments:
-            for p in points:
-                if int(cmt.get("point_id", -1)) == p.id:
-                    code_start_line = cmt.get("code_start_line")
-                    code_end_line = cmt.get("code_end_line")
-                    code = get_code_block_from_patch(patch, code_start_line, code_end_line)
+            try:
+                if cmt.get("commented_file").endswith(".py"):
+                    points = [p for p in points if p.language == "Python"]
+                elif cmt.get("commented_file").endswith(".java"):
+                    points = [p for p in points if p.language == "Java"]
+                else:
+                    continue
+                for p in points:
+                    point_id = int(cmt.get("point_id", -1))
+                    if point_id == p.id:
+                        code_start_line = cmt.get("code_start_line")
+                        code_end_line = cmt.get("code_end_line")
+                        code = get_code_block_from_patch(patch, code_start_line, code_end_line)
 
-                    new_comments.append(
-                        {
-                            "commented_file": cmt.get("commented_file"),
-                            "code": code,
-                            "code_start_line": code_start_line,
-                            "code_end_line": code_end_line,
-                            "comment": cmt.get("comment"),
-                            "point_id": p.id,
-                            "point": p.text,
-                            "point_detail": p.detail,
-                        }
-                    )
-                    break
+                        new_comments.append(
+                            {
+                                "commented_file": cmt.get("commented_file"),
+                                "code": code,
+                                "code_start_line": code_start_line,
+                                "code_end_line": code_end_line,
+                                "comment": cmt.get("comment"),
+                                "point_id": p.id,
+                                "point": p.text,
+                                "point_detail": p.detail,
+                            }
+                        )
+                        break
+            except Exception:
+                pass
 
         logger.debug(f"new_comments: {new_comments}")
         return new_comments
@@ -132,51 +144,49 @@ class CodeReview(Action):
                 code = get_code_block_from_patch(
                     patch, str(max(1, int(code_start_line) - 5)), str(int(code_end_line) + 5)
                 )
+            code_language = "Java"
+            code_file_ext = cmt.get("commented_file", ".java").split(".")[-1]
+            if code_file_ext == ".java":
+                code_language = "Java"
+            elif code_file_ext == ".py":
+                code_language = "Python"
             prompt = CODE_REVIEW_COMFIRM_TEMPLATE.format(
                 code=code,
                 comment=cmt.get("comment"),
                 desc=point.text,
                 example=point.yes_example + "\n" + point.no_example,
             )
-            resp = await self.llm.aask(prompt, system_msgs=[CODE_REVIEW_COMFIRM_SYSTEM_PROMPT])
+            system_prompt = [CODE_REVIEW_COMFIRM_SYSTEM_PROMPT.format(code_language=code_language)]
+            resp = await self.llm.aask(prompt, system_msgs=system_prompt)
             if "True" in resp or "true" in resp:
                 new_comments.append(cmt)
         logger.info(f"original comments num: {len(comments)}, confirmed comments num: {len(new_comments)}")
         return new_comments
 
-    async def cr_by_full_points(self, patch: PatchSet, points: list[Point]):
+    async def cr_by_points(self, patch: PatchSet, points: list[Point]):
         comments = []
-        points_str = "\n".join([f"{p.id} {p.text}" for p in points])
         for patched_file in patch:
+            if not patched_file:
+                continue
             if patched_file.path.endswith(".py"):
-                points_str = "\n".join([f"{p.id} {p.text}" for p in points if p.language == "Python"])
+                points = [p for p in points if p.language == "Python"]
             elif patched_file.path.endswith(".java"):
-                points_str = "\n".join([f"{p.id} {p.text}" for p in points if p.language == "Java"])
+                points = [p for p in points if p.language == "Java"]
             else:
                 continue
-            if len(str(patched_file).splitlines()) >= 50:
-                cr_by_segment_points_comments = await self.cr_by_segment_points(
-                    patched_file=patched_file, points=points
-                )
-                comments += cr_by_segment_points_comments
-                continue
-            prompt = CODE_REVIEW_PROMPT_TEMPLATE.format(patch=str(patched_file), points=points_str)
-            resp = await self.llm.aask(prompt)
-            json_str = parse_json_code_block(resp)[0]
-            comments += json.loads(json_str)
-
-        return comments
-
-    async def cr_by_segment_points(self, patched_file: PatchedFile, points: list[Point]):
-        comments = []
-        group_points = [points[i : i + 3] for i in range(0, len(points), 3)]
-        for group_point in group_points:
-            points_str = "\n".join([f"{p.id} {p.text}" for p in group_point])
-            prompt = CODE_REVIEW_PROMPT_TEMPLATE.format(patch=str(patched_file), points=points_str)
-            resp = await self.llm.aask(prompt)
-            json_str = parse_json_code_block(resp)[0]
-            comments_batch = json.loads(json_str)
-            comments += comments_batch
+            group_points = [points[i : i + 3] for i in range(0, len(points), 3)]
+            for group_point in group_points:
+                points_str = "id description\n"
+                points_str += "\n".join([f"{p.id} {p.text}" for p in group_point])
+                prompt = CODE_REVIEW_PROMPT_TEMPLATE.format(patch=str(patched_file), points=points_str)
+                resp = await self.llm.aask(prompt)
+                json_str = parse_json_code_block(resp)[0]
+                comments_batch = json.loads(json_str)
+                if comments_batch:
+                    patched_file_path = patched_file.path
+                    for c in comments_batch:
+                        c["commented_file"] = patched_file_path
+                    comments += comments_batch
 
         return comments
 
@@ -185,7 +195,7 @@ class CodeReview(Action):
         patch: PatchSet = add_line_num_on_patch(patch)
 
         result = []
-        comments = await self.cr_by_full_points(patch=patch, points=points)
+        comments = await self.cr_by_points(patch=patch, points=points)
         if len(comments) != 0:
             comments = self.format_comments(comments, points, patch)
             comments = await self.confirm_comments(patch=patch, comments=comments, points=points)

metagpt/ext/cr/actions/modify_code.py

@@ -81,17 +81,18 @@ class ModifyCode(Action):
         }
         resp = None
         for patched_file in patch:
-            patch_target_file_name = str(patched_file.target_file).split("/", maxsplit=1)[-1]
-            if patch_target_file_name not in grouped_comments:
+            patch_target_file_name = str(patched_file.path).split("/")[-1]
+            if patched_file.path not in grouped_comments:
                 continue
             comments_prompt = ""
             index = 1
-            for grouped_comment in grouped_comments[patch_target_file_name]:
+            for grouped_comment in grouped_comments[patched_file.path]:
                 comments_prompt += f"""
                     <comment{index}>
                     {grouped_comment}
                     </comment{index}>\n
                 """
+                index += 1
             prompt = MODIFY_CODE_PROMPT.format(patch=patched_file, comments=comments_prompt)
             output_dir = (
                 Path(output_dir)

metagpt/prompts/di/role_zero.py

@@ -10,7 +10,7 @@ Note:
 4. Don't forget to append task first when all existing tasks are finished and new tasks are required.
 5. Don't repeat the same command if it has been executed successfully.
 """
-
+# To ensure compatibility with hard-coded experience, do not add any other content between "# Example" and "# Available Commands".
 CMD_PROMPT = """
 # Data Structure
 class Task(BaseModel):
@@ -19,7 +19,7 @@ class Task(BaseModel):
     instruction: str = ""
     task_type: str = ""
     assignee: str = ""
-
+    
 # Available Task Types
 {task_type_desc}
 
@@ -56,6 +56,7 @@ Some text indicating your thoughts, such as how you should update the plan statu
 ```
 Notice: your output JSON data section must start with **```json [**
 """
+
 JSON_REPAIR_PROMPT = """
 ## json data
 {json_data}
@@ -66,3 +67,10 @@ Formatted JSON data
 ```
 Help check if there are any formatting issues with the JSON data? If so, please help format it
 """
+
+QUICK_THINK_PROMPT = """
+Decide if the latest user message is a quick question.
+Quick questions include common-sense, logical, math questions, greetings, or casual chat that you can answer directly, excluding software development tasks.
+Respond with "#YES#, (then start your actual response to the question...)" if so, otherwise, simply respond with "#NO#".
+Your response:
+"""

metagpt/prompts/di/swe_agent.py

@@ -7,7 +7,7 @@ https://github.com/princeton-nlp/SWE-agent/tree/main/config/configs
 SWE_AGENT_SYSTEM_TEMPLATE = """
 SETTING: You are an autonomous programmer, and you're working directly in the environment line with a special interface.
 
-The special interface consists of a file editor that shows you {WINDOW} lines of a file at a time.
+The special interface consists of a file editor that shows you 100 lines of a file at a time.
 
 Please note that THE EDIT COMMAND REQUIRES PROPER INDENTATION. Pay attention to the original indentation when replacing the function. 
 If you'd like to add the line '        print(x)' you must fully write that out, with all those spaces before the code! Indentation is important and code that is not indented correctly will fail and require fixing before it can be run.
@@ -50,7 +50,8 @@ MINIMAL_EXAMPLE = """
 ## Example of a actions trajectory
 User Requirement and Issue: Fix the bug in the repo. Because the environment is not available, you DO NOT need to run and modify any existing test case files or add new test case files to ensure that the bug is fixed.
 
-### Read and understand issue(Require):
+### Read and understand issue:
+Thought: Firstly, I need to review the detailed information of this issue in order to understand the problem that needs fixing.
 {{
     "command_name": "Browser.goto",
     "args": {{
@@ -60,43 +61,99 @@ User Requirement and Issue: Fix the bug in the repo. Because the environment is
 ->
 
 ### Locate issue(Require): Locate the issue in the code by searching for the relevant file, function, or class and open the file to view the code.
+Thought: I need to come under the repo path
 {{
     "command_name": "Bash.run",
     "args": {{
-        "cmd": "cd /workspace/django__django_3.0"
+        "cmd": "cd /workspace/MetaGPT"
     }}
 }}
 ->
 
-Bash.run(cmd='search_dir_and_preview ASCIIUsernameValidator')
+Thought: Let's start by locating the `openai_api.py` file.\nFirst, let's search for the `openai_api.py` file.
 {{
     "command_name": "Bash.run",
     "args": {{
-        "cmd": "open /workspace/django__django_3.0/django/contrib/auth/validators.py"   
+        "cmd": "find_file 'openai_api.py'"   
+    }}
+}}
+->
+
+Thought: We have located both the `openai_api.py` file. Let's start by opening the `openai_api.py` file to apply the necessary changes.",
+{{
+    "command_name": "Bash.run",
+    "args": {{
+        "cmd": "open '/workspace/MetaGPT/provider/openai_api.py'"   
     }}
 }}
 ->
 
 ### Fix the Bug(Require): Fix the bug in the code by editing the relevant function, class or code snippet.
+Thought: Now that I've found the bug, let's fix it by edit.
 {{
     "command_name": "Bash.run",
     "args": {{
-        "cmd": "edit 10:20 <<EOF\n    regex = r'\A[\w.@+-]+\Z'\n    message = _( \n        'Enter a valid username. This value may contain only English letters, ' \n        'numbers, and @/./+/-/_ characters.'\n    )\n    flags = re.ASCII\n\n@deconstructible\nclass UnicodeUsernameValidator(validators.RegexValidator):\n    regex = r'\A[\w.@+-]+\Z'\nEOF"
+        "cmd": "edit 93:95 <<EOF\n        usage = None\n        collected_messages = []\n        async for chunk in response:\n            if chunk.usage is not None:\n                usage = CompletionUsage(**chunk.usage)\n            chunk_message = chunk.choices[0].delta.content or '' if chunk.choices else ''  # extract the message\n            finish_reason = (\n                chunk.choices[0].finish_reason if chunk.choices and hasattr(chunk.choices[0], 'finish_reason') else None\n            )\n            log_llm_stream(chunk_message)\nEOF"
+    }}
+}}
+->
+Thought: Due to a syntax error related to an undefined name 'Image', we need to address this issue even though it is not directly related to our work. Let's try importing the package to fix it.
+{{
+    "command_name": "Bash.run",
+    "args": {{
+        "cmd": "edit 14:14 <<EOF\nfrom PIL.Image import Image\nEOF"
     }}
 }}
 ->
 
-### Submit the Changes(Require): Submit the changes to the repository.
+### Save the Changes (Required): After all changes have been made, save them to the repository.
+> You must choose one of the following two methods.
+
+#### Just save the changes locally, it only need one action.
+Thought: The bug has been fixed. Let's submit the changes.
 {{
     "command_name": "Bash.run",
     "args": {{
         "cmd": "submit"
     }}
 }}
-Bash.run(cmd='submit')
 ->
+
+#### Save the changes and commit them to the remote repository.
+
+##### Push the changes from the local repository to the remote repository.
+Thought: All changes have been saved, let's push the code to the remote repository.
 {{
-    "command_name": "end",
+    "command_name": "git_push",
+    "args": {{
+        "local_path": "/workspace/MetaGPT",
+        "app_name": "github",
+        "comments": "Fix Issue #1275: produced TypeError: openai.types.completion_usage.CompletionUsage() argument after ** must be a mapping, not NoneType"",
+        "new_branch": "test-fix"
+    }}
+}}
+->
+
+##### Create a pull request (Optional): Merge the changes from the new branch into the master branch.
+Thought: Now that the changes have been pushed to the remote repository, due to the user's requirement, let's create a pull request to merge the changes into the master branch.
+[{{
+    "command_name": "git_create_pull",
+    "args": {{
+        "base": "master",
+        "head": "test-fix",
+        "base_repo_name": "garylin2099/MetaGPT",
+        "head_repo_name": "seeker-jie/MetaGPT",
+        "app_name": "github",
+        "title": "Fix Issue #1275: produced TypeError: openai.types.completion_usage.CompletionUsage() argument after ** must be a mapping, not NoneType"",
+        "body": "This pull request addresses issue #1275 by ensuring that chunk.usage is not None before passing it to CompletionUsage."
+   }}
+}}]
+->
+
+### Finally
+Thought: All task has been done, let's end the conversation.
+{{
+    "command_name": "end"
 }}
 """
 
@@ -162,7 +219,10 @@ IMPORTANT_TIPS = """
     - If a search command fails, modify the search criteria and check for typos or incorrect paths, then try again.
     - Based on feedback of observation or bash command in trajectory to guide adjustments in your search strategy.
 
-13. If the task results in succeed, fail, or NO PROGRESS, output `submit`.
+13. Save the code change:
+  - If you need to submit changes to the remote repository, first use the regular git commit command to save the changes locally, then select a command from the `Available Commands: [git_push, git_create_pull]` to submit the changes to the remote repository.
+
+  - If you don't need to submit code changes to the remote repository. use the command Bash.run('submit') to commit the changes locally. 
 
 14. If provided an issue link, you MUST go to the issue page using Browser tool to understand the issue before starting your fix.
 
@@ -185,5 +245,4 @@ The current bash state is:
 (Current directory: {{working_dir}})
 
 Avoid repeating the same command. Instead, please think about the current situation and provide the next bash command to execute in JSON format:"
-
 """

metagpt/roles/di/role_zero.py

@@ -8,12 +8,16 @@ from typing import Callable, Dict, List, Literal, Tuple
 
 from pydantic import model_validator
 
-from metagpt.actions import Action
+from metagpt.actions import Action, UserRequirement
 from metagpt.actions.di.run_command import RunCommand
+from metagpt.exp_pool import exp_cache
+from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
+from metagpt.exp_pool.serializers import RoleZeroSerializer
 from metagpt.logs import logger
 from metagpt.prompts.di.role_zero import (
     CMD_PROMPT,
     JSON_REPAIR_PROMPT,
+    QUICK_THINK_PROMPT,
     ROLE_INSTRUCTION,
 )
 from metagpt.roles import Role
@@ -24,7 +28,7 @@ from metagpt.tools.libs.browser import Browser
 from metagpt.tools.libs.editor import Editor
 from metagpt.tools.tool_recommend import BM25ToolRecommender, ToolRecommender
 from metagpt.tools.tool_registry import register_tool
-from metagpt.utils.common import CodeParser
+from metagpt.utils.common import CodeParser, any_to_str
 from metagpt.utils.repair_llm_raw_output import RepairType, repair_llm_raw_output
 from metagpt.utils.report import ThoughtReporter
 
@@ -142,24 +146,42 @@ class RoleZero(Role):
         tool_info = json.dumps({tool.name: tool.schemas for tool in tools})
 
         ### Make Decision Dynamically ###
+        instruction = self.instruction.strip()
         prompt = self.cmd_prompt.format(
-            plan_status=plan_status,
-            current_task=current_task,
             example=example,
             available_commands=tool_info,
-            instruction=self.instruction.strip(),
             task_type_desc=self.task_type_desc,
+            plan_status=plan_status,
+            current_task=current_task,
+            instruction=instruction,
         )
         memory = self.rc.memory.get(self.memory_k)
         memory = await self.parse_browser_actions(memory)
-        context = self.llm.format_msg(memory + [UserMessage(content=prompt)])
-        # print(*context, sep="\n" + "*" * 5 + "\n")
-        async with ThoughtReporter(enable_llm_stream=True):
-            self.command_rsp = await self.llm.aask(context, system_msgs=self.system_msg)
+
+        req = self.llm.format_msg(memory + [UserMessage(content=prompt)])
+        async with ThoughtReporter(enable_llm_stream=True) as reporter:
+            await reporter.async_report({"type": "react"})
+            state_data = dict(
+                plan_status=plan_status,
+                current_task=current_task,
+                instruction=instruction,
+            )
+            self.command_rsp = await self.llm_cached_aask(req=req, system_msgs=self.system_msg, state_data=state_data)
+
         self.rc.memory.add(AIMessage(content=self.command_rsp))
 
         return True
 
+    @exp_cache(context_builder=RoleZeroContextBuilder(), serializer=RoleZeroSerializer())
+    async def llm_cached_aask(self, *, req: list[dict], system_msgs: list[str], **kwargs) -> str:
+        """Use `exp_cache` to automatically manage experiences.
+
+        The `RoleZeroContextBuilder` attempts to add experiences to `req`.
+        The `RoleZeroSerializer` extracts essential parts of `req` for the experience pool, trimming lengthy entries to retain only necessary parts.
+        """
+
+        return await self.llm.aask(req, system_msgs=system_msgs)
+
     async def parse_browser_actions(self, memory: List[Message]) -> List[Message]:
         if not self.browser.is_empty_page:
             pattern = re.compile(r"Command Browser\.(\w+) executed")
@@ -189,9 +211,14 @@ class RoleZero(Role):
         )
 
     async def _react(self) -> Message:
-        # NOTE: Diff 1: Each time landing here means observing news, set todo to allow news processing in _think
+        # NOTE: Diff 1: Each time landing here means news is observed, set todo to allow news processing in _think
         self._set_state(0)
 
+        # problems solvable by quick thinking doesn't need to a formal think-act cycle
+        quick_rsp = await self._quick_think()
+        if quick_rsp:
+            return quick_rsp
+
         actions_taken = 0
         rsp = AIMessage(content="No actions taken yet", cause_by=Action)  # will be overwritten after Role _act
         while actions_taken < self.rc.max_react_loop:
@@ -208,6 +235,31 @@ class RoleZero(Role):
             actions_taken += 1
         return rsp  # return output from the last action
 
+    async def _quick_think(self) -> Message:
+        msg = self.rc.news[-1]
+        rsp_msg = None
+        if msg.cause_by != any_to_str(UserRequirement):
+            # Agents themselves won't generate quick questions, use this rule to reduce extra llm calls
+            return rsp_msg
+
+        context = self.llm.format_msg(self.get_memories(k=4) + [UserMessage(content=QUICK_THINK_PROMPT)])
+        async with ThoughtReporter(enable_llm_stream=True) as reporter:
+            await reporter.async_report({"type": "quick"})
+            rsp = await self.llm.aask(context)
+
+        pattern = r"#YES#,? ?"
+        if re.search(pattern, rsp):
+            answer = re.sub(pattern, "", rsp).strip()
+            self.rc.memory.add(AIMessage(content=answer, cause_by=RunCommand))
+            await self.reply_to_human(content=answer)
+            rsp_msg = AIMessage(
+                content="Complete run",
+                sent_from=self.name,
+                cause_by=RunCommand,
+            )
+
+        return rsp_msg
+
     async def _parse_commands(self) -> Tuple[List[Dict], bool]:
         """Retrieves commands from the Large Language Model (LLM).
 
@@ -223,6 +275,7 @@ class RoleZero(Role):
             commands = CodeParser.parse_code(block=None, lang="json", text=self.command_rsp)
             commands = json.loads(repair_llm_raw_output(output=commands, req_keys=[None], repair_type=RepairType.JSON))
         except json.JSONDecodeError:
+            logger.warning(f"Failed to parse JSON for: {self.command_rsp}. Trying to repair...")
             commands = await self.llm.aask(msg=JSON_REPAIR_PROMPT.format(json_data=self.command_rsp))
             commands = json.loads(CodeParser.parse_code(block=None, lang="json", text=commands))
         except Exception as e:

metagpt/roles/di/swe_agent.py

@@ -1,5 +1,4 @@
 import json
-import os
 
 from pydantic import Field
 
@@ -18,9 +17,7 @@ class SWEAgent(RoleZero):
     name: str = "Swen"
     profile: str = "Issue Solver"
     goal: str = "Resolve GitHub issue"
-    _bash_window_size: int = 100
-    _system_msg: str = SWE_AGENT_SYSTEM_TEMPLATE
-    system_msg: list[str] = [_system_msg.format(WINDOW=_bash_window_size)]
+    system_msg: str = [SWE_AGENT_SYSTEM_TEMPLATE]
     _instruction: str = NEXT_STEP_TEMPLATE
     tools: list[str] = [
         "Bash",
@@ -35,8 +32,7 @@ class SWEAgent(RoleZero):
     run_eval: bool = False
 
     async def _think(self) -> bool:
-        self._update_system_msg()
-        self._format_instruction()
+        await self._format_instruction()
         res = await super()._think()
         if self.run_eval:
             await self._parse_commands_for_eval()
@@ -51,32 +47,16 @@ class SWEAgent(RoleZero):
             }
         )
 
-    def _update_system_msg(self):
-        """
-        Sets the system message for the SWE agent.
-
-        Sets the `_bash_window_size` from the environment variable `WINDOW` if it exists.
-        Formats the `_system_msg` template with the current `_bash_window_size`.
-        """
-        if os.getenv("WINDOW"):
-            self._bash_window_size = int(os.getenv("WINDOW"))
-        self.system_msg = [self._system_msg.format(WINDOW=self._bash_window_size)]
-
-    def _format_instruction(self):
+    async def _format_instruction(self):
         """
         Formats the instruction message for the SWE agent.
 
         Runs the "state" command in the terminal, parses its output as JSON,
         and uses it to format the `_instruction` template.
         """
-        state_output = self.terminal.run("state")
+        state_output = await self.terminal.run("state")
         bash_state = json.loads(state_output)
-
-        self.instruction = self._instruction.format(
-            WINDOW=self._bash_window_size, examples=MINIMAL_EXAMPLE, **bash_state
-        ).strip()
-
-        return self.instruction
+        self.instruction = self._instruction.format(**bash_state).strip()
 
     async def _parse_commands_for_eval(self):
         """
@@ -97,7 +77,7 @@ class SWEAgent(RoleZero):
             if "end" != cmd.get("command_name", ""):
                 return
         try:
-            diff_output = self.terminal.run("git diff --cached")
+            diff_output = await self.terminal.run("git diff --cached")
             clear_diff = extract_patch(diff_output)
             logger.info(f"Diff output: \n{clear_diff}")
             if clear_diff:

metagpt/schema.py

@@ -669,14 +669,7 @@ class Plan(BaseModel):
         """
         return [task for task in self.tasks if task.is_finished]
 
-    def append_task(
-        self,
-        task_id: str,
-        dependent_task_ids: list[str],
-        instruction: str,
-        assignee: str,
-        task_type: str = ""
-    ):
+    def append_task(self, task_id: str, dependent_task_ids: list[str], instruction: str, assignee: str, task_type: str = ""):
         """Append a new task with task_id (number) to the end of existing task sequences. If dependent_task_ids is not empty, the task will depend on the tasks with the ids in the list."""
         new_task = Task(
             task_id=task_id,

metagpt/strategy/experience_retriever.py

@@ -510,15 +510,6 @@ Explanation: The requirement is about software development. Assign each tasks to
             "assignee": "Alex"
         }
     },
-    {
-        "command_name": "Plan.append_task",
-        "args": {
-            "task_id": "5",
-            "dependent_task_ids": ["4"],
-            "instruction": "Write comprehensive tests for the game logic and user interface to ensure functionality and reliability.",
-            "assignee": "Edward"
-        }
-    },
     {
         "command_name": "TeamLeader.publish_message",
         "args": {
@@ -802,13 +793,13 @@ Explanation: I will first need to read the system design document and the projec
     {
         "command_name": "Editor.read",
         "args": {
-            "path": "/tmp/docs/project_schedule.json"
+            "path": "/tmp/project_schedule.json"
         }
     },
     {
         "command_name": "Editor.read",
         "args": {
-            "path": "/tmp/docs/system_design.json"
+            "path": "/tmp/system_design.json"
         }
     }
 ]

metagpt/strategy/solver.py

@@ -39,7 +39,7 @@ class NaiveSolver(BaseSolver):
         self.graph.topological_sort()
         for key in self.graph.execution_order:
             op = self.graph.nodes[key]
-            await op.fill(self.context, self.llm, mode="root")
+            await op.fill(req=self.context, llm=self.llm, mode="root")
 
 
 class TOTSolver(BaseSolver):

metagpt/tools/libs/browser.py

@@ -24,6 +24,7 @@ from metagpt.utils.a11y_tree import (
     scroll_page,
     type_text,
 )
+from metagpt.utils.proxy_env import get_proxy_from_env
 from metagpt.utils.report import BrowserReporter
 
 
@@ -72,7 +73,7 @@ class Browser:
         self.page: Optional[Page] = None
         self.accessibility_tree: list = []
         self.headless: bool = True
-        self.proxy = None
+        self.proxy = get_proxy_from_env()
         self.is_empty_page = True
         self.reporter = BrowserReporter()
 
@@ -120,7 +121,7 @@ class Browser:
         await scroll_page(self.page, direction)
         return await self._wait_page()
 
-    async def goto(self, url: str, timeout: float = 30000):
+    async def goto(self, url: str, timeout: float = 90000):
         """Navigate to a specific URL."""
         if self.page is None:
             await self.start()

metagpt/tools/libs/cr.py

@@ -3,6 +3,7 @@ from pathlib import Path
 from typing import Optional
 
 import aiofiles
+from bs4 import BeautifulSoup
 from unidiff import PatchSet
 
 import metagpt.ext.cr
@@ -29,7 +30,7 @@ class CodeReview:
         Args:
             patch_path: The local path of the patch file or the url of the pull request. Example: "/data/xxx-pr-1.patch", "https://github.com/xx/XX/pull/1362"
             cr_output_file: Output file path where code review comments will be saved. Example: "cr/xxx-pr-1.json"
-            cr_point_file: File path for specifying code review points. Defaults to a predefined file.
+            cr_point_file: File path for specifying code review points. If not specified, this parameter is not passed..
         """
         patch = await self._get_patch_content(patch_path)
         cr_point_file = cr_point_file if cr_point_file else Path(metagpt.ext.cr.__file__).parent / "points.json"
@@ -45,7 +46,7 @@ class CodeReview:
             )
             comments = await CodeReview_().run(patch, cr_points)
             cr_output_path.parent.mkdir(exist_ok=True, parents=True)
-            async with aiofiles.open(cr_output_path, "w") as f:
+            async with aiofiles.open(cr_output_path, "w", encoding="utf-8") as f:
                 await f.write(json.dumps(comments, ensure_ascii=False))
             await reporter.async_report(cr_output_path)
 
@@ -65,7 +66,7 @@ class CodeReview:
             output_dir: File path where code review comments are stored.
         """
         patch = await self._get_patch_content(patch_path)
-        async with aiofiles.open(cr_file, "r") as f:
+        async with aiofiles.open(cr_file, "r", encoding="utf-8") as f:
             comments = json.loads(await f.read())
         await ModifyCode(pr="").run(patch, comments, output_dir)
         return f"The fixed patch files store in {output_dir}"
@@ -75,12 +76,14 @@ class CodeReview:
             # async with aiohttp.ClientSession(trust_env=True) as client:
             #     async with client.get(f"{patch_path}.diff", ) as resp:
             #         patch_file_content = await resp.text()
-            browser = Browser()
-            browser.proxy = {"server": "http://127.0.0.1:20172"}
-            async with browser:
+            async with Browser() as browser:
                 await browser.goto(f"{patch_path}.diff")
                 patch_file_content = await browser.page.content()
-
+                if patch_file_content.startswith("<html>"):
+                    soup = BeautifulSoup(patch_file_content, "html.parser")
+                    pre = soup.find("pre")
+                    if pre:
+                        patch_file_content = pre.text
         else:
             async with aiofiles.open(patch_path) as f:
                 patch_file_content = await f.read()

metagpt/tools/libs/git.py

@@ -11,7 +11,7 @@ from github.PullRequest import PullRequest
 from metagpt.tools.tool_registry import register_tool
 
 
-@register_tool(tags=["software development", "git", "Commit the changes and push to remote git repository."])
+@register_tool(tags=["software development", "git", "Push to remote git repository."])
 async def git_push(
     local_path: Union[str, Path],
     app_name: str,
@@ -22,22 +22,23 @@ async def git_push(
     Pushes changes from a local Git repository to its remote counterpart.
 
     Args:
-        local_path (Union[str, Path]): The path to the local Git repository.
-        app_name (str): The name of the application where the repository is hosted. For example, "github", "gitlab", "bitbucket", etc.
-        comments (str, optional): The commit message to use. Defaults to "Commit".
+        local_path (Union[str, Path]): The absolute path to the local Git repository.
+        app_name (str): The name of the platform hosting the repository (e.g., "github", "gitlab", "bitbucket").
+        comments (str, optional): Comments to be associated with the push. Defaults to "Commit".
         new_branch (str, optional): The name of the new branch to create and push changes to.
             If not provided, changes will be pushed to the current branch. Defaults to "".
 
     Returns:
         GitBranch: The branch to which the changes were pushed.
+
     Raises:
         ValueError: If the provided local_path does not point to a valid Git repository.
 
     Example:
         >>> url = "https://github.com/iorisa/snake-game.git"
         >>> local_path = await git_clone(url=url)
-        >>> app_name="github"
-        >>> comments = "Archive"
+        >>> app_name = "github"
+        >>> comments = "Commit"
         >>> new_branch = "feature/new"
         >>> branch = await git_push(local_path=local_path, app_name=app_name, comments=comments, new_branch=new_branch)
         >>> base = branch.base
@@ -45,8 +46,8 @@ async def git_push(
         >>> repo_name = branch.repo_name
         >>> print(f"base branch:'{base}', head branch:'{head}', repo_name:'{repo_name}'")
         base branch:'master', head branch:'feature/new', repo_name:'iorisa/snake-game'
-
     """
+
     from metagpt.tools.libs import get_env
     from metagpt.utils.git_repository import GitRepository
 
@@ -64,9 +65,9 @@ async def git_push(
 async def git_create_pull(
     base: str,
     head: str,
-    base_repo_name: str,
     app_name: str,
-    head_repo_name: Optional[str] = None,
+    base_repo_name: str,
+    head_repo_name: str = None,
     title: Optional[str] = None,
     body: Optional[str] = None,
     issue: Optional[Issue] = None,
@@ -75,14 +76,14 @@ async def git_create_pull(
     Creates a pull request on a Git repository. Use this tool in priority over Browser to create a pull request.
 
     Args:
-        base (str): The base branch of the pull request.
-        head (str): The head branch of the pull request.
-        base_repo_name (str): The full repository name (user/repo) where the pull request will be created.
-        app_name (str): The name of the application where the repository is hosted. For example, "github", "gitlab", "bitbucket", etc.
-        head_repo_name (Optional[str], optional): The full repository name (user/repo) where the pull request will merge from. Defaults to None.
-        title (Optional[str], optional): The title of the pull request. Defaults to None.
-        body (Optional[str], optional): The body of the pull request. Defaults to None.
-        issue (Optional[Issue], optional): The related issue of the pull request. Defaults to None.
+        base (str): The name of the base branch where the pull request will be merged.
+        head (str): The name of the branch that contains the changes for the pull request.
+        app_name (str): The name of the platform hosting the repository (e.g., "github", "gitlab", "bitbucket").
+        base_repo_name (str): The full name of the target repository (in the format "user/repo") where the pull request will be created.
+        head_repo_name (Optional[str]): The full name of the source repository (in the format "user/repo") from which the changes will be pulled.
+        title (Optional[str]): The title of the pull request. Defaults to None.
+        body (Optional[str]): The description or body content of the pull request. Defaults to None.
+        issue (Optional[Issue]): An optional issue related to the pull request. Defaults to None.
 
     Example:
         >>> # create pull request

metagpt/tools/libs/terminal.py

@@ -1,8 +1,10 @@
-import subprocess
-import threading
-from queue import Queue
+import asyncio
+from asyncio import Queue
+from asyncio.subprocess import PIPE, STDOUT
+from typing import Optional
 
 from metagpt.const import DEFAULT_WORKSPACE_ROOT, SWE_SETUP_PATH
+from metagpt.logs import logger
 from metagpt.tools.tool_registry import register_tool
 from metagpt.utils.report import END_MARKER_VALUE, TerminalReporter
 
@@ -19,62 +21,54 @@ class Terminal:
     def __init__(self):
         self.shell_command = ["bash"]  # FIXME: should consider windows support later
         self.command_terminator = "\n"
-
-        # Start a persistent shell process
-        self.process = subprocess.Popen(
-            self.shell_command,
-            shell=True,
-            stdin=subprocess.PIPE,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.STDOUT,
-            executable="/bin/bash",
-        )
-        self.stdout_queue = Queue()
+        self.stdout_queue = Queue(maxsize=1000)
         self.observer = TerminalReporter()
+        self.process: Optional[asyncio.subprocess.Process] = None
 
-        self._check_state()
+    async def _start_process(self):
+        # Start a persistent shell process
+        self.process = await asyncio.create_subprocess_exec(
+            *self.shell_command, stdin=PIPE, stdout=PIPE, stderr=STDOUT, executable="bash"
+        )
+        await self._check_state()
 
-    def _check_state(self):
-        """Check the state of the terminal, e.g. the current directory of the terminal process. Useful for agent to understand."""
-        print("The terminal is at:", self.run_command("pwd"))
+    async def _check_state(self):
+        """
+        Check the state of the terminal, e.g. the current directory of the terminal process. Useful for agent to understand.
+        """
+        output = await self.run_command("pwd")
+        logger.info("The terminal is at:", output)
 
-    def run_command(self, cmd: str, daemon=False) -> str:
+    async def run_command(self, cmd: str, daemon=False) -> str:
         """
         Executes a specified command in the terminal and streams the output back in real time.
         This command maintains state across executions, such as the current directory,
-        allowing for sequential commands to be contextually aware. The output from the
-        command execution is placed into `stdout_queue`, which can be consumed as needed.
+        allowing for sequential commands to be contextually aware.
 
         Args:
             cmd (str): The command to execute in the terminal.
-            daemon (bool): If True, executes the command in a background thread, allowing
-                           the main program to continue execution. The command's output is
-                           collected asynchronously in daemon mode and placed into `stdout_queue`.
-
+            daemon (bool): If True, executes the command in an asynchronous task, allowing
+                           the main program to continue execution.
         Returns:
             str: The command's output or an empty string if `daemon` is True. Remember that
-                 when `daemon` is True, the output is collected into `stdout_queue` and must
-                 be consumed from there.
-
-        Note:
-            If `stdout_queue` is not periodically consumed, it could potentially grow indefinitely,
-            consuming memory. Ensure that there's a mechanism in place to consume this queue,
-            especially during long-running or output-heavy command executions.
+                 when `daemon` is True, use the `get_stdout_output` method to get the output.
         """
+        if self.process is None:
+            await self._start_process()
 
         # Send the command
         self.process.stdin.write((cmd + self.command_terminator).encode())
         self.process.stdin.write(
-            (f'echo "{END_MARKER_VALUE}"{self.command_terminator}').encode()  # write EOF
+            f'echo "{END_MARKER_VALUE}"{self.command_terminator}'.encode()  # write EOF
         )  # Unique marker to signal command end
-        self.process.stdin.flush()
+        await self.process.stdin.drain()
         if daemon:
-            threading.Thread(target=self._read_and_process_output, args=(cmd,), daemon=True).start()
+            asyncio.create_task(self._read_and_process_output(cmd))
             return ""
         else:
-            return self._read_and_process_output(cmd)
+            return await self._read_and_process_output(cmd)
 
-    def execute_in_conda_env(self, cmd: str, env, daemon=False) -> str:
+    async def execute_in_conda_env(self, cmd: str, env, daemon=False) -> str:
         """
         Executes a given command within a specified Conda environment automatically without
         the need for manual activation. Users just need to provide the name of the Conda
@@ -84,7 +78,7 @@ class Terminal:
             cmd (str): The command to execute within the Conda environment.
             env (str, optional): The name of the Conda environment to activate before executing the command.
                                  If not specified, the command will run in the current active environment.
-            daemon (bool): If True, the command is run in a background thread, similar to `run_command`,
+            daemon (bool): If True, the command is run in an asynchronous task, similar to `run_command`,
                            affecting error logging and handling in the same manner.
 
         Returns:
@@ -96,19 +90,34 @@ class Terminal:
             to ensure the specified environment is active for the command's execution.
         """
         cmd = f"conda run -n {env} {cmd}"
-        return self.run_command(cmd, daemon=daemon)
+        return await self.run_command(cmd, daemon=daemon)
 
-    def _read_and_process_output(self, cmd):
-        with self.observer as observer:
+    async def get_stdout_output(self) -> str:
+        """
+        Retrieves all collected output from background running commands and returns it as a string.
+
+        Returns:
+            str: The collected output from background running commands, returned as a string.
+        """
+        output_lines = []
+        while not self.stdout_queue.empty():
+            line = await self.stdout_queue.get()
+            output_lines.append(line)
+        return "\n".join(output_lines)
+
+    async def _read_and_process_output(self, cmd, daemon=False) -> str:
+        async with self.observer as observer:
             cmd_output = []
-            observer.report(cmd + self.command_terminator, "cmd")
-            # report the comman
+            await observer.async_report(cmd + self.command_terminator, "cmd")
+            # report the command
             # Read the output until the unique marker is found.
             # We read bytes directly from stdout instead of text because when reading text,
             # '\r' is changed to '\n', resulting in excessive output.
             tmp = b""
             while True:
-                output = tmp + self.process.stdout.read(1)
+                output = tmp + await self.process.stdout.read(1)
+                if not output:
+                    continue
                 *lines, tmp = output.splitlines(True)
                 for line in lines:
                     line = line.decode()
@@ -116,20 +125,20 @@ class Terminal:
                     if ix >= 0:
                         line = line[0:ix]
                         if line:
-                            observer.report(line, "output")
+                            await observer.async_report(line, "output")
                             # report stdout in real-time
                             cmd_output.append(line)
                         return "".join(cmd_output)
                     # log stdout in real-time
-                    observer.report(line, "output")
+                    await observer.async_report(line, "output")
                     cmd_output.append(line)
-                    self.stdout_queue.put(line)
+                    if daemon:
+                        await self.stdout_queue.put(line)
 
-    def close(self):
+    async def close(self):
         """Close the persistent shell process."""
         self.process.stdin.close()
-        self.process.terminate()
-        self.process.wait()
+        await self.process.wait()
 
 
 @register_tool(include_functions=["run"])
@@ -142,10 +151,13 @@ class Bash(Terminal):
     def __init__(self):
         """init"""
         super().__init__()
-        self.run_command(f"cd {DEFAULT_WORKSPACE_ROOT}")
-        self.run_command(f"source {SWE_SETUP_PATH}")
+        self.start_flag = False
 
-    def run(self, cmd) -> str:
+    async def start(self):
+        await self.run_command(f"cd {DEFAULT_WORKSPACE_ROOT}")
+        await self.run_command(f"source {SWE_SETUP_PATH}")
+
+    async def run(self, cmd) -> str:
         """
         Executes a bash command.
 
@@ -184,9 +196,6 @@ class Bash(Terminal):
           Arguments:
               filename (str): The name of the file to create.
 
-        - submit
-          Submits your current code. it can only be executed once, the last action before the `end`.
-
         - search_dir_and_preview <search_term> [<dir>]
           Searches for search_term in all files in dir and gives their code preview
           with line numbers. If dir is not provided, searches in the current directory.
@@ -220,6 +229,13 @@ class Bash(Terminal):
               end_line (int): The line number to end the edit at (inclusive), starting from 1.
               replacement_text (str): The text to replace the current selection with, must conform to PEP8 standards.
 
+        - submit
+          Submits your current code locally. it can only be executed once, the last action before the `end`.
+
         Note: Make sure to use these functions as per their defined arguments and behaviors.
         """
-        return self.run_command(cmd)
+        if not self.start_flag:
+            await self.start()
+            self.start_flag = True
+
+        return await self.run_command(cmd)

metagpt/tools/swe_agent_commands/setup_default.sh

@@ -17,4 +17,3 @@ source $REPO_ROOT_DIR/metagpt/tools/swe_agent_commands/search.sh
 source $REPO_ROOT_DIR/metagpt/tools/swe_agent_commands/edit_linting.sh
 
 export SWE_CMD_WORK_DIR="$REPO_ROOT_DIR/workspace/swe_agent_workdir"
-#sudo chmod 777 $REPO_ROOT_DIR/workspace/swe_agent_workdir

metagpt/utils/proxy_env.py

@@ -0,0 +1,19 @@
+import os
+
+
+def get_proxy_from_env():
+    proxy_config = {}
+    server = None
+    for i in ("ALL_PROXY", "all_proxy", "HTTPS_PROXY", "https_proxy", "HTTP_PROXY", "http_proxy"):
+        if os.environ.get(i):
+            server = os.environ.get(i)
+    if server:
+        proxy_config["server"] = server
+    no_proxy = os.environ.get("NO_PROXY") or os.environ.get("no_proxy")
+    if no_proxy:
+        proxy_config["bypass"] = no_proxy
+
+    if not proxy_config:
+        proxy_config = None
+
+    return proxy_config

metagpt/utils/report.py

@@ -266,10 +266,6 @@ class ThoughtReporter(ObjectReporter):
 
     block: Literal[BlockType.THOUGHT] = BlockType.THOUGHT
 
-    async def __aenter__(self):
-        await self.async_report({})
-        return await super().__aenter__()
-
 
 class FileReporter(ResourceReporter):
     """File resource callback for reporting complete file paths.

metagpt/utils/token_counter.py

@@ -31,9 +31,9 @@ TOKEN_COSTS = {
     "gpt-4-0125-preview": {"prompt": 0.01, "completion": 0.03},
     "gpt-4-1106-preview": {"prompt": 0.01, "completion": 0.03},
     "gpt-4-vision-preview": {"prompt": 0.01, "completion": 0.03},  # TODO add extra image price calculator
-    "gpt-4-1106-vision-preview": {"prompt": 0.01, "completion": 0.03},
     "gpt-4o": {"prompt": 0.005, "completion": 0.015},
     "gpt-4o-2024-05-13": {"prompt": 0.005, "completion": 0.015},
+    "gpt-4-1106-vision-preview": {"prompt": 0.01, "completion": 0.03},
     "text-embedding-ada-002": {"prompt": 0.0004, "completion": 0.0},
     "glm-3-turbo": {"prompt": 0.0007, "completion": 0.0007},  # 128k version, prompt + completion tokens=0.005￥/k-tokens
     "glm-4": {"prompt": 0.014, "completion": 0.014},  # 128k version, prompt + completion tokens=0.1￥/k-tokens
@@ -147,11 +147,14 @@ FIREWORKS_GRADE_TOKEN_COSTS = {
 
 # https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo
 TOKEN_MAX = {
+    "gpt-4o-2024-05-13": 128000,
+    "gpt-4o": 128000,
     "gpt-4-0125-preview": 128000,
     "gpt-4-turbo-preview": 128000,
     "gpt-4-1106-preview": 128000,
     "gpt-4-vision-preview": 128000,
     "gpt-4-1106-vision-preview": 128000,
+    "gpt-4-turbo": 128000,
     "gpt-4": 8192,
     "gpt-4-0613": 8192,
     "gpt-4-32k": 32768,