Skip to content

Build and run agents you can see, understand and trust.

19.6 KB
Raw Blame
# -*- coding: utf-8 -*-
"""Test the goal pipeline."""
from types import SimpleNamespace
from typing import Any, AsyncGenerator
from unittest.async_case import IsolatedAsyncioTestCase

from utils import AnyString

from agentscope.event import (
    ConfirmResult,
    RequireUserConfirmEvent,
    UserConfirmResultEvent,
)
from agentscope.message import (
    Base64Source,
    DataBlock,
    Msg,
    TextBlock,
    ToolCallBlock,
    UserMsg,
)
from agentscope.pipeline import GoalPipeline
from agentscope.types import ReplyFinishedReason


def _report(text: str = "已完成,见 main.py") -> Msg:
    """An executor's final message carrying its achievement report."""
    return Msg(
        name="executor",
        content=[],
        role="assistant",
        finished_reason=ReplyFinishedReason.COMPLETED,
        structured_output={"report": text},
    )


def _verdict(result: str, message: str = "") -> Msg:
    """A verifier's final message carrying a structured verdict."""
    return Msg(
        name="verifier",
        content=[],
        role="assistant",
        finished_reason=ReplyFinishedReason.COMPLETED,
        structured_output={"result": result, "message": message},
    )


def _no_output(name: str) -> Msg:
    """A final message from an agent that never called the output tool."""
    return Msg(
        name=name,
        content=[],
        role="assistant",
        finished_reason=ReplyFinishedReason.COMPLETED,
    )


def _confirm_request() -> RequireUserConfirmEvent:
    """A tool call parked on a human."""
    return RequireUserConfirmEvent(
        reply_id="executor-reply",
        tool_calls=[
            ToolCallBlock(id="call-1", name="write_file", input="{}"),
        ],
    )


class StubAgent:
    """Replays one scripted batch of chunks per ``reply_stream`` call.

    Records what it was handed, so a test can assert the feedback and
    reminders the pipeline built reached the right agent.
    """

    def __init__(self, name: str, script: list[list[Any]]) -> None:
        """Initialize the stub with one script entry per call."""
        self.name = name
        self.script = script
        self.state = SimpleNamespace(
            reply_id=f"{name}-reply",
            context=[],
            summary="",
        )
        self.received: list[Any] = []
        self.conversation_before_calls: list[tuple[list[Any], Any]] = []

    # pylint: disable=unused-argument
    async def reply_stream(
        self,
        inputs: Any = None,
        structured_schema: Any = None,
        yield_final_msg: bool = False,
    ) -> AsyncGenerator[Any, None]:
        """Yield the next batch, holding the final message back unless it
        was asked for, the way ``Agent.reply_stream`` does."""
        self.received.append(inputs)
        self.conversation_before_calls.append(
            (list(self.state.context), self.state.summary),
        )
        batch = self.script[min(len(self.received) - 1, len(self.script) - 1)]
        for chunk in batch:
            if isinstance(chunk, Msg) and not yield_final_msg:
                continue
            yield chunk


class GoalPipelineTest(IsolatedAsyncioTestCase):
    """The goal pipeline test case."""

    async def asyncSetUp(self) -> None:
        """Prepare the input every test starts from."""
        self.query = UserMsg(name="user", content="写一个爬虫")

    async def _run(self, pipe: GoalPipeline, inputs: Any) -> list:
        """Drain one pipeline run into what it yielded."""
        return [chunk async for chunk in pipe.reply_stream(inputs)]

    async def test_passes_on_first_round(self) -> None:
        """A passing verdict ends the run, and the verifier is told both
        the goal and what the executor reported."""
        executor = StubAgent("executor", [[_report("见 main.py")]])
        verifier = StubAgent("verifier", [[_verdict("pass")]])
        pipe = GoalPipeline(executor, verifier)

        yielded = await self._run(pipe, self.query)

        self.assertListEqual(
            [chunk.structured_output for chunk in yielded],
            [{"report": "见 main.py"}],
        )
        told = verifier.received[0].get_text_content()
        self.assertIn("写一个爬虫", told)
        self.assertIn("见 main.py", told)

    async def test_refusal_reaches_the_executor(self) -> None:
        """A refusal is fed back verbatim and the run tries again."""
        executor = StubAgent("executor", [[_report()], [_report()]])
        verifier = StubAgent(
            "verifier",
            [[_verdict("fail", "缺 requirements.txt")], [_verdict("pass")]],
        )
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, self.query)

        self.assertEqual(len(executor.received), 2)
        self.assertIn(
            "缺 requirements.txt",
            executor.received[1].get_text_content(),
        )

    async def test_resets_verifier_conversation_between_refused_rounds(
        self,
    ) -> None:
        """The default keeps a refused verdict out of the next check."""
        executor = StubAgent("executor", [[_report()], [_report()]])
        verifier = StubAgent(
            "verifier",
            [[_verdict("fail", "try again")], [_verdict("pass")]],
        )
        prior = UserMsg("user", "prior verdict")
        verifier.state.context.append(prior)
        verifier.state.summary = "prior summary"
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, self.query)

        self.assertListEqual(
            verifier.conversation_before_calls,
            [([prior], "prior summary"), ([], "")],
        )

    async def test_can_keep_verifier_conversation_between_refused_rounds(
        self,
    ) -> None:
        """Opting out preserves the verifier's existing conversation."""
        executor = StubAgent("executor", [[_report()], [_report()]])
        verifier = StubAgent(
            "verifier",
            [[_verdict("fail", "try again")], [_verdict("pass")]],
        )
        prior = UserMsg("user", "prior verdict")
        verifier.state.context.append(prior)
        verifier.state.summary = "prior summary"
        pipe = GoalPipeline(executor, verifier, verifier_reset_context=False)

        await self._run(pipe, self.query)

        self.assertListEqual(
            verifier.conversation_before_calls,
            [([prior], "prior summary"), ([prior], "prior summary")],
        )

    async def test_stops_at_max_iters(self) -> None:
        """A verdict that never passes stops once the budget is spent."""
        executor = StubAgent("executor", [[_report()]])
        verifier = StubAgent("verifier", [[_verdict("fail", "还是不行")]])
        pipe = GoalPipeline(executor, verifier, max_iters=2)

        await self._run(pipe, self.query)

        self.assertEqual(len(executor.received), 2)
        self.assertEqual(len(verifier.received), 2)

    async def test_impossible_ends_the_run(self) -> None:
        """An impossible goal settles the run rather than retrying."""
        executor = StubAgent("executor", [[_report()]])
        verifier = StubAgent(
            "verifier",
            [[_verdict("impossible", "目标自相矛盾")]],
        )
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, self.query)

        self.assertEqual(len(executor.received), 1)
        self.assertEqual(len(verifier.received), 1)

    async def test_reprompts_a_verifier_that_skips_the_tool(self) -> None:
        """A final message with no verdict is not a refusal: the verifier
        is reminded rather than the executor being sent back."""
        executor = StubAgent("executor", [[_report()]])
        verifier = StubAgent(
            "verifier",
            [[_no_output("verifier")], [_verdict("pass")]],
        )
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, self.query)

        self.assertEqual(len(verifier.received), 2)
        self.assertIn(
            "GenerateStructuredOutput",
            verifier.received[1].get_text_content(),
        )
        # A malfunction is not charged to the executor.
        self.assertEqual(len(executor.received), 1)

    async def test_retry_keeps_a_multimodal_goal(self) -> None:
        """The retry reminder carries a multimodal goal as blocks."""
        query = UserMsg(
            name="user",
            content=[
                TextBlock(text="Check this chart"),
                DataBlock(
                    source=Base64Source(
                        data="aGVsbG8=",
                        media_type="image/png",
                    ),
                ),
            ],
        )
        executor = StubAgent("executor", [[_report()]])
        verifier = StubAgent(
            "verifier",
            [[_no_output("verifier")], [_verdict("pass")]],
        )
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, query)

        self.assertListEqual(
            [block.model_dump() for block in verifier.received[1].content],
            [
                {
                    "type": "text",
                    "text": "<system-reminder>You have failed to generate "
                    "valid verification result. You should call the "
                    "'GenerateStructuredOutput' tool with a valid structured "
                    "output that matches the schema. Recall the verification "
                    "requirements as follows:\n<goal>",
                    "id": AnyString(),
                    "created_at": AnyString(),
                    "finished_at": None,
                },
                {
                    "type": "text",
                    "text": "Check this chart",
                    "id": AnyString(),
                    "created_at": AnyString(),
                    "finished_at": None,
                },
                {
                    "type": "data",
                    "id": AnyString(),
                    "source": {
                        "type": "base64",
                        "data": "aGVsbG8=",
                        "media_type": "image/png",
                    },
                    "name": None,
                    "created_at": AnyString(),
                    "finished_at": None,
                },
                {
                    "type": "text",
                    "text": "</goal></system-reminder>",
                    "id": AnyString(),
                    "created_at": AnyString(),
                    "finished_at": None,
                },
            ],
        )

    async def test_reprompts_an_executor_that_skips_the_tool(self) -> None:
        """The same for the executor: a missing report is asked for again
        rather than read through as if it were there."""
        executor = StubAgent(
            "executor",
            [[_no_output("executor")], [_report()]],
        )
        verifier = StubAgent("verifier", [[_verdict("pass")]])
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, self.query)

        self.assertEqual(len(executor.received), 2)
        self.assertIn(
            "GenerateStructuredOutput",
            executor.received[1].get_text_content(),
        )
        self.assertEqual(len(verifier.received), 1)

    async def test_gives_up_on_an_executor_that_never_reports(self) -> None:
        """Stop after the initial attempt and the allowed retries."""
        for max_retries in (0, 2):
            with self.subTest(max_retries=max_retries):
                missing = _no_output("executor")
                executor = StubAgent("executor", [[missing]])
                verifier = StubAgent("verifier", [[_verdict("pass")]])
                pipe = GoalPipeline(
                    executor,
                    verifier,
                    max_retries=max_retries,
                )
                yielded = []

                with self.assertRaises(RuntimeError) as raised:
                    async for chunk in pipe.reply_stream(
                        UserMsg("user", "Build a crawler"),
                    ):
                        yielded.append(chunk)

                self.assertEqual(
                    str(raised.exception),
                    "The executor failed to generate a valid execution "
                    f"report after {max_retries} retries.",
                )
                self.assertEqual(yielded, [missing] * (max_retries + 1))
                self.assertEqual(len(executor.received), max_retries + 1)
                self.assertEqual(verifier.received, [])

    async def test_gives_up_on_a_verifier_that_never_verdicts(self) -> None:
        """Bound retries with a missing verdict or an empty stream."""
        for batch in ([], [_no_output("verifier")]):
            with self.subTest(empty_stream=not batch):
                report = _report("Completed; see main.py")
                executor = StubAgent("executor", [[report]])
                verifier = StubAgent("verifier", [batch])
                pipe = GoalPipeline(executor, verifier, max_retries=2)
                yielded = []

                with self.assertRaises(RuntimeError) as raised:
                    async for chunk in pipe.reply_stream(
                        UserMsg("user", "Build a crawler"),
                    ):
                        yielded.append(chunk)

                self.assertEqual(
                    str(raised.exception),
                    "The verifier failed to generate a valid verification "
                    "result after 2 retries.",
                )
                self.assertEqual(yielded, [report])
                self.assertEqual(
                    (len(executor.received), len(verifier.received)),
                    (1, 3),
                )

    async def test_executor_can_retry_in_each_round(self) -> None:
        """Each report can use the full retry budget."""
        missing = _no_output("executor")
        first, second = _report("v1"), _report("v2")
        executor = StubAgent(
            "executor",
            [[missing], [first], [missing], [second]],
        )
        verifier = StubAgent(
            "verifier",
            [[_verdict("fail", "fix it")], [_verdict("pass")]],
        )
        pipe = GoalPipeline(executor, verifier, max_retries=1)

        yielded = await self._run(pipe, self.query)

        self.assertEqual(yielded, [missing, first, missing, second])
        self.assertEqual(
            (len(executor.received), len(verifier.received)),
            (4, 2),
        )

    async def test_verifier_can_retry_in_each_round(self) -> None:
        """Each verdict can use the full retry budget."""
        first, second = _report("v1"), _report("v2")
        executor = StubAgent("executor", [[first], [second]])
        verifier = StubAgent(
            "verifier",
            [[], [_verdict("fail", "fix it")], [], [_verdict("pass")]],
        )
        pipe = GoalPipeline(executor, verifier, max_retries=1)

        yielded = await self._run(pipe, self.query)

        self.assertEqual(yielded, [first, second])
        self.assertEqual(
            (len(executor.received), len(verifier.received)),
            (2, 4),
        )

    async def test_a_fresh_run_restarts_the_retry_budget(self) -> None:
        """Each participant can retry again after a previous run failed."""
        for participant in ("executor", "verifier"):
            with self.subTest(participant=participant):
                report = _report("Completed; see main.py")
                verdict = _verdict("pass")
                executor = StubAgent("executor", [[report]])
                verifier = StubAgent("verifier", [[verdict]])
                missing = _no_output(participant)
                if participant == "executor":
                    executor.script = [[missing]] * 3 + [[report]]
                    expected_yielded = [missing, report]
                    expected_calls = (4, 1)
                    output_name = "execution report"
                else:
                    verifier.script = [[missing]] * 3 + [[verdict]]
                    expected_yielded = [report]
                    expected_calls = (2, 4)
                    output_name = "verification result"
                pipe = GoalPipeline(executor, verifier, max_retries=1)

                with self.assertRaises(RuntimeError) as raised:
                    await self._run(
                        pipe,
                        UserMsg("user", "Build a crawler"),
                    )
                self.assertEqual(
                    str(raised.exception),
                    f"The {participant} failed to generate a valid "
                    f"{output_name} after 1 retries.",
                )

                # The new run must retry once before succeeding.
                yielded = await self._run(
                    pipe,
                    UserMsg("user", "Build another crawler"),
                )
                self.assertEqual(yielded, expected_yielded)
                self.assertEqual(
                    (len(executor.received), len(verifier.received)),
                    expected_calls,
                )

    async def test_a_parked_executor_is_not_verified(self) -> None:
        """The work is unfinished while the executor waits on a human, so
        there is nothing for the verifier to judge yet."""
        request = _confirm_request()
        executor = StubAgent("executor", [[request]])
        verifier = StubAgent("verifier", [[_verdict("pass")]])
        pipe = GoalPipeline(executor, verifier)

        yielded = await self._run(pipe, self.query)

        self.assertListEqual(yielded, [request])
        self.assertListEqual(verifier.received, [])

    async def test_resumes_into_the_agent_that_parked(self) -> None:
        """The reply id sends the answer back to whoever asked for it."""
        request = _confirm_request()
        executor = StubAgent("executor", [[request], [_report()]])
        verifier = StubAgent("verifier", [[_verdict("pass")]])
        pipe = GoalPipeline(executor, verifier)

        await self._run(pipe, self.query)

        answer = UserConfirmResultEvent(
            reply_id="executor-reply",
            confirm_results=[
                ConfirmResult(confirmed=True, tool_call=request.tool_calls[0]),
            ],
        )
        await self._run(pipe, answer)

        self.assertEqual(executor.received[1], answer)
        self.assertEqual(len(verifier.received), 1)

    async def test_resume_keeps_the_iteration_budget(self) -> None:
        """Resuming does not hand the run a fresh set of attempts.

        Round one is refused and round two parks. With a budget of two,
        the resumed round is the last one — were the budget to restart,
        the executor would be sent back a fourth time.
        """
        executor = StubAgent(
            "executor",
            [[_report()], [_confirm_request()], [_report()]],
        )
        verifier = StubAgent(
            "verifier",
            [[_verdict("fail", "不行")], [_verdict("fail", "还是不行")]],
        )
        pipe = GoalPipeline(executor, verifier, max_iters=2)

        await self._run(pipe, self.query)
        await self._run(
            pipe,
            UserConfirmResultEvent(
                reply_id="executor-reply",
                confirm_results=[],
            ),
        )

        self.assertEqual(len(executor.received), 3)
        self.assertEqual(len(verifier.received), 2)

    async def test_rejects_an_unknown_reply_id(self) -> None:
        """An answer belonging to neither agent is a programming error,
        not something to guess at."""
        executor = StubAgent("executor", [[_report()]])
        verifier = StubAgent("verifier", [[_verdict("pass")]])
        pipe = GoalPipeline(executor, verifier)

        with self.assertRaises(ValueError):
            await self._run(
                pipe,
                UserConfirmResultEvent(
                    reply_id="nobody",
                    confirm_results=[],
                ),
            )