core/agents/bug_hunter.py - main - githubmirror/gpt-pilot

core/agents/bug_hunter.py
411 строк19 KB
Zvonimir Sabljic
Merge branch 'v2' into v2_new
21 июн 2025, 06:58
21 июн 2025, 06:58687504e
import asyncio
import json
from enum import Enum

from pydantic import BaseModel, Field

from core.agents.base import BaseAgent
from core.agents.convo import AgentConvo
from core.agents.mixins import ChatWithBreakdownMixin, TestSteps
from core.agents.response import AgentResponse
from core.config import CHECK_LOGS_AGENT_NAME, magic_words
from core.config.actions import (
    BH_ADDITIONAL_FEEDBACK,
    BH_HUMAN_TEST_AGAIN,
    BH_IS_BUG_FIXED,
    BH_START_BUG_HUNT,
    BH_START_USER_TEST,
    BH_STARTING_PAIR_PROGRAMMING,
    BH_WAIT_BUG_REP_INSTRUCTIONS,
)
from core.config.constants import CONVO_ITERATIONS_LIMIT
from core.db.models.project_state import IterationStatus
from core.llm.parser import JSONParser
from core.log import get_logger
from core.telemetry import telemetry
from core.ui.base import ProjectStage, pythagora_source

log = get_logger(__name__)


class HuntConclusionType(str, Enum):
    ADD_LOGS = magic_words.ADD_LOGS
    PROBLEM_IDENTIFIED = magic_words.PROBLEM_IDENTIFIED


class HuntConclusionOptions(BaseModel):
    conclusion: HuntConclusionType = Field(
        description=f"If more logs are needed to identify the problem, respond with '{magic_words.ADD_LOGS}'. If the problem is identified, respond with '{magic_words.PROBLEM_IDENTIFIED}'."
    )


class ImportantLog(BaseModel):
    logCode: str = Field(description="Actual line of code that prints the log.")
    shouldBeDifferent: bool = Field(
        description="Whether the current output should be different from the expected output."
    )
    filePath: str = Field(description="Path to the file in which the log exists.")
    currentOutput: str = Field(description="Current output of the log.")
    expectedOutput: str = Field(description="Expected output of the log.")
    explanation: str = Field(description="A brief explanation of the log.")


class ImportantLogsForDebugging(BaseModel):
    logs: list[ImportantLog] = Field(description="Important logs that will help the human debug the current bug.")


class BugHunter(ChatWithBreakdownMixin, BaseAgent):
    agent_type = "bug-hunter"
    display_name = "Bug Hunter"

    async def run(self) -> AgentResponse:
        current_iteration = self.current_state.current_iteration

        if "bug_reproduction_description" not in current_iteration:
            if not self.state_manager.async_tasks:
                self.state_manager.async_tasks = []
                self.state_manager.async_tasks.append(asyncio.create_task(self.get_bug_reproduction_instructions()))

        if current_iteration["status"] == IterationStatus.HUNTING_FOR_BUG:
            # TODO determine how to find a bug (eg. check in db, ask user a question, etc.)
            return await self.check_logs()
        elif current_iteration["status"] == IterationStatus.AWAITING_USER_TEST:
            await self.ui.send_bug_hunter_status("close_status", 0)
            return await self.ask_user_to_test(False, True)
        elif current_iteration["status"] == IterationStatus.AWAITING_BUG_REPRODUCTION:
            await self.ui.send_bug_hunter_status("close_status", 0)
            return await self.ask_user_to_test(True, False)
        elif current_iteration["status"] == IterationStatus.START_PAIR_PROGRAMMING:
            await self.ui.send_bug_hunter_status("close_status", 0)
            return await self.start_pair_programming()

    async def get_bug_reproduction_instructions(self):
        await self.send_message("Finding a way to reproduce the bug ...")
        await self.ui.set_important_stream()
        llm = self.get_llm()
        convo = (
            AgentConvo(self)
            .template(
                "get_bug_reproduction_instructions",
                current_task=self.current_state.current_task,
                user_feedback=self.current_state.current_iteration["user_feedback"],
                user_feedback_qa=self.current_state.current_iteration["user_feedback_qa"],
                docs=self.current_state.docs,
                next_solution_to_try=None,
            )
            .require_schema(TestSteps)
        )
        bug_reproduction_instructions: TestSteps = await llm(convo, parser=JSONParser(TestSteps), temperature=0)
        self.next_state.current_iteration["bug_reproduction_description"] = json.dumps(
            [test.dict() for test in bug_reproduction_instructions.steps]
        )

    async def check_logs(self, logs_message: str = None):
        self.next_state.action = BH_START_BUG_HUNT.format(
            self.current_state.tasks.index(self.current_state.current_task) + 1
        )
        llm = self.get_llm(CHECK_LOGS_AGENT_NAME, stream_output=True)
        convo = self.generate_iteration_convo_so_far()
        await self.ui.start_breakdown_stream()
        human_readable_instructions = await llm(convo, temperature=0.5)

        convo.assistant(human_readable_instructions)

        human_readable_instructions = await self.chat_with_breakdown(convo, human_readable_instructions)

        convo = (
            AgentConvo(self)
            .template(
                "bug_found_or_add_logs",
                hunt_conclusion=human_readable_instructions,
            )
            .require_schema(HuntConclusionOptions)
        )
        llm = self.get_llm()
        hunt_conclusion = await llm(convo, parser=JSONParser(HuntConclusionOptions), temperature=0)

        bug_hunting_cycles = self.current_state.current_iteration.get("bug_hunting_cycles")
        num_bug_hunting_cycles = len(bug_hunting_cycles) if bug_hunting_cycles else 0
        if hunt_conclusion.conclusion == magic_words.PROBLEM_IDENTIFIED:
            # if no need for logs, implement iteration same as before
            self.set_data_for_next_hunting_cycle(human_readable_instructions, IterationStatus.AWAITING_BUG_FIX)
            await self.send_message("Found the bug. I'm attempting to fix it ...")
            await self.ui.send_bug_hunter_status("fixing_bug", num_bug_hunting_cycles)
        else:
            # if logs are needed, add logging steps
            self.set_data_for_next_hunting_cycle(human_readable_instructions, IterationStatus.AWAITING_LOGGING)
            await self.send_message("Adding more logs to identify the bug ...")
            await self.ui.send_bug_hunter_status("adding_logs", num_bug_hunting_cycles)

        self.next_state.flag_iterations_as_modified()
        await self.async_task_finish()
        return AgentResponse.done(self)

    async def ask_user_to_test(self, awaiting_bug_reproduction: bool = False, awaiting_user_test: bool = False):
        if awaiting_user_test:
            self.next_state.action = BH_START_USER_TEST.format(
                self.current_state.tasks.index(self.current_state.current_task) + 1
            )
        elif awaiting_bug_reproduction:
            self.next_state.action = BH_WAIT_BUG_REP_INSTRUCTIONS.format(
                self.current_state.tasks.index(self.current_state.current_task) + 1
            )

        await self.async_task_finish()

        test_instructions = self.current_state.current_iteration["bug_reproduction_description"]
        await self.ui.send_message(
            "Start the app and test it by following these instructions:\n\n", source=pythagora_source
        )
        await self.ui.send_test_instructions(test_instructions, project_state_id=str(self.current_state.id))

        if self.current_state.run_command:
            await self.ui.send_run_command(self.current_state.run_command)

        user_feedback = await self.ask_question(
            BH_HUMAN_TEST_AGAIN,
            buttons={"done": "I am done testing"},
            buttons_only=True,
            default="continue",
            extra_info={"restart_app": True},
            hint="Instructions for testing:\n\n" + test_instructions,
        )

        if awaiting_user_test:
            self.next_state.current_iteration["bug_hunting_cycles"][-1]["fix_attempted"] = True

        if awaiting_user_test and not user_feedback.text:
            buttons = {"yes": "Yes, the issue is fixed", "no": "No", "start_pair_programming": "Start Pair Programming"}
            user_feedback = await self.ask_question(
                BH_IS_BUG_FIXED,
                buttons=buttons,
                default="yes",
                buttons_only=True,
                hint="Instructions for testing:\n\n" + test_instructions,
            )
            # self.next_state.current_iteration["bug_hunting_cycles"][-1]["fix_attempted"] = True

            if user_feedback.button == "yes":
                self.next_state.complete_iteration()
                return AgentResponse.done(self)
            elif user_feedback.button == "start_pair_programming":
                self.next_state.current_iteration["status"] = IterationStatus.START_PAIR_PROGRAMMING
                self.next_state.flag_iterations_as_modified()
                return AgentResponse.done(self)
            else:
                awaiting_bug_reproduction = True

        if awaiting_bug_reproduction and not user_feedback.text:
            buttons = {
                "done": "Bug is fixed",
                "continue": "Continue without feedback",  # DO NOT CHANGE THIS TEXT without changing it in the extension (it is hardcoded)
                "start_pair_programming": "Start Pair Programming",
            }
            await self.ui.send_project_stage(
                {
                    "stage": ProjectStage.ADDITIONAL_FEEDBACK,
                }
            )
            user_feedback = await self.ask_question(
                BH_ADDITIONAL_FEEDBACK,
                buttons=buttons,
                default="continue",
                extra_info={"collect_logs": True},
                hint="Instructions for testing:\n\n" + test_instructions,
            )

            if user_feedback.button == "done":
                self.next_state.complete_iteration()
                return AgentResponse.done(self)
            elif user_feedback.button == "start_pair_programming":
                self.next_state.current_iteration["status"] = IterationStatus.START_PAIR_PROGRAMMING
                self.next_state.flag_iterations_as_modified()
                return AgentResponse.done(self)

        # TODO select only the logs that are new (with PYTHAGORA_DEBUGGING_LOG)
        self.next_state.current_iteration["bug_hunting_cycles"][-1]["backend_logs"] = None
        self.next_state.current_iteration["bug_hunting_cycles"][-1]["frontend_logs"] = None
        self.next_state.current_iteration["bug_hunting_cycles"][-1]["user_feedback"] = user_feedback.text
        self.next_state.current_iteration["status"] = IterationStatus.HUNTING_FOR_BUG
        self.next_state.current_iteration["attempts"] += 1
        self.next_state.flag_iterations_as_modified()

        await self.ui.send_project_stage(
            {
                "bug_fix_attempt": self.next_state.current_iteration["attempts"],
            }
        )

        return AgentResponse.done(self)

    async def start_pair_programming(self):
        self.next_state.action = BH_STARTING_PAIR_PROGRAMMING.format(
            self.current_state.tasks.index(self.current_state.current_task) + 1
        )
        llm = self.get_llm(stream_output=True)
        convo = self.generate_iteration_convo_so_far(True)
        if len(convo.messages) > 1:
            convo.remove_last_x_messages(1)
        convo = convo.template("problem_explanation")
        await self.ui.set_important_stream()
        initial_explanation = await llm(convo, temperature=0.5)

        llm = self.get_llm()
        convo = convo.template("data_about_logs").require_schema(ImportantLogsForDebugging)
        data_about_logs = await llm(convo, parser=JSONParser(ImportantLogsForDebugging), temperature=0.5)

        await self.ui.send_data_about_logs(
            {
                "logs": [
                    {
                        "currentLog": d.currentOutput,
                        "expectedLog": d.expectedOutput,
                        "explanation": d.explanation,
                        "filePath": d.filePath,
                        "logCode": d.logCode,
                        "shouldBeDifferent": d.shouldBeDifferent,
                    }
                    for d in data_about_logs.logs
                ]
            }
        )

        await self.async_task_finish()

        while True:
            self.next_state.current_iteration["initial_explanation"] = initial_explanation
            next_step = await self.ask_question(
                "What do you want to do?",
                buttons={
                    "question": "I have a question",
                    "done": "I fixed the bug myself",
                    "tell_me_more": "Tell me more about the bug",
                    "solution_hint": "I think I know where the problem is",
                    "other": "Other",
                },
                buttons_only=True,
                default="continue",
                hint="Instructions for testing:\n\n"
                + self.current_state.current_iteration["bug_reproduction_description"],
            )

            await telemetry.trace_code_event(
                "pair-programming",
                {
                    "button": next_step.button,
                    "num_tasks": len(self.current_state.tasks),
                    "num_epics": len(self.current_state.epics),
                    "num_iterations": len(self.current_state.iterations),
                    "app_id": str(self.state_manager.project.id),
                    "app_name": self.state_manager.project.name,
                    "folder_name": self.state_manager.project.folder_name,
                },
            )

            # TODO: remove when Leon checks
            convo.remove_last_x_messages(2)

            if len(convo.messages) > CONVO_ITERATIONS_LIMIT:
                convo.slice(1, CONVO_ITERATIONS_LIMIT)

            # TODO: in the future improve with a separate conversation that parses the user info and goes into an appropriate if statement
            if next_step.button == "done":
                self.next_state.complete_iteration()
                break
            elif next_step.button == "question":
                user_response = await self.ask_question("Oh, cool, what would you like to know?")
                convo = convo.template("ask_a_question", question=user_response.text)
                await self.ui.set_important_stream()
                llm_answer = await llm(convo, temperature=0.5)
                await self.send_message(llm_answer)
            elif next_step.button == "tell_me_more":
                convo.template("tell_me_more")
                await self.ui.set_important_stream()
                response = await llm(convo, temperature=0.5)
                await self.send_message(response)
            elif next_step.button == "other":
                # this is the same as "question" - we want to keep an option for users to click to understand if we're missing something with other options
                user_response = await self.ask_question("Let me know what you think ...")
                convo = convo.template("ask_a_question", question=user_response.text)
                await self.ui.set_important_stream()
                llm_answer = await llm(convo, temperature=0.5)
                await self.send_message(llm_answer)
            elif next_step.button == "solution_hint":
                human_hint_label = "Amazing! How do you think we can solve this bug?"
                while True:
                    human_hint = await self.ask_question(human_hint_label)
                    convo = convo.template("instructions_from_human_hint", human_hint=human_hint.text)
                    await self.ui.set_important_stream()
                    llm = self.get_llm(CHECK_LOGS_AGENT_NAME, stream_output=True)
                    human_readable_instructions = await llm(convo, temperature=0.5)
                    human_approval = await self.ask_question(
                        "Can I implement this solution?", buttons={"yes": "Yes", "no": "No"}, buttons_only=True
                    )
                    llm = self.get_llm(stream_output=True)
                    if human_approval.button == "yes":
                        self.set_data_for_next_hunting_cycle(
                            human_readable_instructions, IterationStatus.AWAITING_BUG_FIX
                        )
                        self.next_state.flag_iterations_as_modified()
                        break
                    else:
                        human_hint_label = "Oh, my bad, what did I misunderstand?"
                break
            elif next_step.button == "tell_me_more":
                convo.template("tell_me_more")
                await self.ui.set_important_stream()
                response = await llm(convo, temperature=0.5)
                await self.send_message(response)
                continue

        return AgentResponse.done(self)

    def generate_iteration_convo_so_far(self, omit_last_cycle=False):
        convo = AgentConvo(self).template(
            "iteration",
            current_task=self.current_state.current_task,
            user_feedback=self.current_state.current_iteration["user_feedback"],
            user_feedback_qa=self.current_state.current_iteration["user_feedback_qa"],
            docs=self.current_state.docs,
            magic_words=magic_words,
            next_solution_to_try=None,
            test_instructions=json.loads(self.current_state.current_task.get("test_instructions") or "[]"),
        )

        hunting_cycles = self.current_state.current_iteration.get("bug_hunting_cycles", [])[
            0 : (-1 if omit_last_cycle else None)
        ]

        for hunting_cycle in hunting_cycles:
            convo = convo.assistant(hunting_cycle["human_readable_instructions"]).template(
                "log_data",
                backend_logs=hunting_cycle.get("backend_logs"),
                frontend_logs=hunting_cycle.get("frontend_logs"),
                fix_attempted=hunting_cycle.get("fix_attempted"),
                user_feedback=hunting_cycle.get("user_feedback"),
            )

        if len(convo.messages) > CONVO_ITERATIONS_LIMIT:
            convo.slice(1, CONVO_ITERATIONS_LIMIT)

        return convo

    async def async_task_finish(self):
        if self.state_manager.async_tasks:
            if not self.state_manager.async_tasks[-1].done():
                await self.send_message("Waiting for the bug reproduction instructions...")
                await self.state_manager.async_tasks[-1]
            self.state_manager.async_tasks = []

    def set_data_for_next_hunting_cycle(self, human_readable_instructions, new_status):
        self.next_state.current_iteration["description"] = human_readable_instructions
        self.next_state.current_iteration["bug_hunting_cycles"] += [
            {
                "human_readable_instructions": human_readable_instructions,
                "fix_attempted": any(
                    c["fix_attempted"] for c in self.current_state.current_iteration["bug_hunting_cycles"]
                ),
            }
        ]

        self.next_state.current_iteration["status"] = new_status