tests/pipeline_engine/test_pipeline_e2e.py

"""End-to-end tests for the pipeline engine.

These tests exercise realistic multi-step scenarios — no ACE imports.
Dummy steps simulate the Agent → Evaluate → Reflect → Update pattern using
a test-local StepContext subclass, so the full plumbing is exercised.
"""

from __future__ import annotations

import asyncio
import threading
import time
from dataclasses import dataclass
from types import MappingProxyType
from typing import Any

import pytest

from pipeline import (
    Branch,
    MergeStrategy,
    Pipeline,
    SampleResult,
    StepContext,
)

# ---------------------------------------------------------------------------
# Test-local context subclass (simulates ACE-style domain fields)
# ---------------------------------------------------------------------------


@dataclass(frozen=True)
class E2EContext(StepContext):
    """Domain context with named fields for the Agent → Evaluate → Reflect → Update pattern."""

    agent_output: Any = None
    environment_result: Any = None
    reflection: Any = None
    skill_manager_output: Any = None


# ---------------------------------------------------------------------------
# Domain-agnostic dummy steps (ACE-shaped but no ACE imports)
# ---------------------------------------------------------------------------


class AgentStep:
    """Reads ctx.sample (a named field, not metadata), writes agent_output."""

    requires = frozenset()
    provides = frozenset({"agent_output"})

    def __call__(self, ctx: E2EContext) -> E2EContext:
        return ctx.replace(agent_output=f"answer_for_{ctx.sample}")


class EvaluateStep:
    """Reads agent_output + environment from metadata, writes environment_result."""

    requires = frozenset({"agent_output"})
    provides = frozenset({"environment_result"})

    def __call__(self, ctx: E2EContext) -> E2EContext:
        correct = ctx.agent_output == ctx.metadata.get("expected")
        return ctx.replace(
            environment_result={
                "correct": correct,
                "feedback": "ok" if correct else "wrong",
            }
        )


class ReflectStep:
    """Background step: reads agent_output + environment_result, writes reflection."""

    requires = frozenset({"agent_output", "environment_result"})
    provides = frozenset({"reflection"})
    async_boundary = True
    max_workers = 3

    def __call__(self, ctx: E2EContext) -> E2EContext:
        time.sleep(0.01)  # simulate LLM latency
        return ctx.replace(
            reflection={
                "insight": "reflected",
                "correct": ctx.environment_result["correct"],
            }
        )


class UpdateStep:
    """Background step: reads reflection, writes skill_manager_output. Serialized."""

    requires = frozenset({"reflection"})
    provides = frozenset({"skill_manager_output"})
    max_workers = 1
    _updates: list = []
    _lock = threading.Lock()

    def __call__(self, ctx: E2EContext) -> E2EContext:
        time.sleep(0.005)
        with self._lock:
            UpdateStep._updates.append(ctx.reflection)
        return ctx.replace(skill_manager_output={"updated": True})


class LogStep:
    """Side-effect step that records sample name (for Branch tests)."""

    requires = frozenset()
    provides = frozenset()

    def __init__(self):
        self.log: list[str] = []
        self._lock = threading.Lock()

    def __call__(self, ctx: StepContext) -> StepContext:
        with self._lock:
            self.log.append(ctx.sample)
        return ctx.replace(metadata=MappingProxyType({**ctx.metadata, "logged": True}))


class MetricStep:
    """Writes a metric to metadata (for Branch tests alongside ReflectStep)."""

    requires = frozenset()
    provides = frozenset()

    def __call__(self, ctx: StepContext) -> StepContext:
        return ctx.replace(
            metadata=MappingProxyType({**ctx.metadata, "metric": len(str(ctx.sample))})
        )


# ---------------------------------------------------------------------------
# E2E: full 4-step pipeline (no async_boundary first)
# ---------------------------------------------------------------------------


@pytest.mark.integration
class TestFullPipelineChain:
    def _pipe(self) -> Pipeline:
        return Pipeline().then(AgentStep()).then(EvaluateStep())

    def test_single_sample_correct_answer(self):
        sample_ctx_kwargs = {"metadata": {"expected": "answer_for_q1"}}
        ctx = E2EContext(sample="q1", **sample_ctx_kwargs)
        # Run via __call__ (nested mode)
        out = self._pipe()(ctx)
        assert out.agent_output == "answer_for_q1"
        assert out.environment_result["correct"] is True

    def test_multiple_samples_run(self):
        # Samples without expected = all "wrong"
        results = self._pipe().run([E2EContext(sample=s) for s in ("q1", "q2", "q3")])
        assert len(results) == 3
        assert all(r.error is None for r in results)
        assert all(r.output.agent_output.startswith("answer_for_") for r in results)

    def test_data_flows_correctly_through_chain(self):
        results = self._pipe().run([E2EContext(sample="hello")])
        out = results[0].output
        assert out.sample == "hello"
        assert out.agent_output == "answer_for_hello"
        assert out.environment_result is not None
        assert "correct" in out.environment_result


# ---------------------------------------------------------------------------
# E2E: async_boundary — fire-and-forget pipeline
# ---------------------------------------------------------------------------


@pytest.mark.integration
@pytest.mark.slow
class TestAsyncBoundaryPipeline:
    def _pipe(self) -> Pipeline:
        UpdateStep._updates.clear()
        return (
            Pipeline()
            .then(AgentStep())
            .then(EvaluateStep())
            .then(ReflectStep())  # async_boundary = True
            .then(UpdateStep())
        )

    def test_run_returns_before_background_finishes(self):
        pipe = self._pipe()
        t0 = time.monotonic()
        results = pipe.run([E2EContext(sample="q1")], workers=1)
        foreground_time = time.monotonic() - t0
        # Should return quickly (foreground only: agent + evaluate)
        # Background runs Reflect (0.01s) + Update (0.005s) asynchronously
        assert len(results) == 1
        pipe.wait_for_background(timeout=5.0)

    def test_all_steps_complete_after_wait(self):
        pipe = self._pipe()
        results = pipe.run([E2EContext(sample="q1"), E2EContext(sample="q2")])
        pipe.wait_for_background(timeout=5.0)
        assert all(r.output is not None for r in results)
        assert all(r.output.skill_manager_output == {"updated": True} for r in results)

    def test_update_serialized_across_samples(self):
        """UpdateStep.max_workers=1 — updates must not interleave."""
        UpdateStep._updates.clear()
        pipe = self._pipe()
        contexts = [E2EContext(sample=f"s{i}") for i in range(5)]
        pipe.run(contexts, workers=3)
        pipe.wait_for_background(timeout=10.0)
        # All 5 samples must have triggered an update
        assert len(UpdateStep._updates) == 5

    def test_failed_foreground_not_sent_to_background(self):
        class FailEval:
            requires = frozenset({"agent_output"})
            provides = frozenset({"environment_result"})

            def __call__(self, ctx):
                raise RuntimeError("eval failed")

        UpdateStep._updates.clear()
        pipe = (
            Pipeline()
            .then(AgentStep())
            .then(FailEval())
            .then(ReflectStep())
            .then(UpdateStep())
        )
        results = pipe.run([E2EContext(sample="q1")])
        pipe.wait_for_background(timeout=2.0)
        # Error in foreground → no background submission
        assert results[0].error is not None
        assert results[0].failed_at == "FailEval"
        assert len(UpdateStep._updates) == 0


# ---------------------------------------------------------------------------
# E2E: Branch inside a pipeline
# ---------------------------------------------------------------------------


@pytest.mark.integration
class TestBranchInPipeline:
    def test_branch_parallel_reflect_and_log(self):
        log_step = LogStep()
        pipe = (
            Pipeline()
            .then(AgentStep())
            .then(EvaluateStep())
            .branch(
                Pipeline().then(MetricStep()),
                Pipeline().then(log_step),
                merge=MergeStrategy.RAISE_ON_CONFLICT,
            )
        )
        results = pipe.run([E2EContext(sample="hello"), E2EContext(sample="world")])
        assert len(results) == 2
        assert all(r.error is None for r in results)
        # Both branches ran
        assert all(r.output.metadata.get("metric") is not None for r in results)
        assert sorted(log_step.log) == ["hello", "world"]

    def test_step_after_branch_receives_merged_context(self):
        class Summarize:
            requires = frozenset()
            provides = frozenset({"summary"})

            def __call__(self, ctx):
                return ctx.replace(
                    metadata=MappingProxyType({**ctx.metadata, "summary": "done"})
                )

        pipe = (
            Pipeline()
            .then(AgentStep())
            .branch(
                Pipeline().then(MetricStep()),
                Pipeline().then(LogStep()),
            )
            .then(Summarize())
        )
        results = pipe.run([E2EContext(sample="test")])
        assert results[0].output.metadata.get("summary") == "done"

    def test_branch_failure_captured_in_sample_result(self):
        class BranchBoom:
            requires = frozenset()
            provides = frozenset()

            def __call__(self, ctx):
                raise RuntimeError("branch_fail")

        pipe = (
            Pipeline()
            .then(AgentStep())
            .branch(
                Pipeline().then(MetricStep()),
                Pipeline().then(BranchBoom()),
            )
        )
        results = pipe.run([E2EContext(sample="s")])
        assert results[0].error is not None
        assert results[0].failed_at == "Branch"


# ---------------------------------------------------------------------------
# E2E: nested pipeline reuse
# ---------------------------------------------------------------------------


@pytest.mark.integration
class TestNestedPipelineReuse:
    def test_inner_pipeline_reused_in_two_outer_pipelines(self):
        inner = Pipeline().then(AgentStep()).then(EvaluateStep())

        outer_a = Pipeline().then(inner)
        outer_b = Pipeline().then(inner).then(MetricStep())

        r_a = outer_a.run([E2EContext(sample="q1")])
        r_b = outer_b.run([E2EContext(sample="q1")])

        assert r_a[0].output.agent_output == "answer_for_q1"
        assert r_b[0].output.metadata.get("metric") is not None

    def test_deeply_nested_pipelines(self):
        level1 = Pipeline().then(AgentStep())
        level2 = Pipeline().then(level1).then(EvaluateStep())
        level3 = Pipeline().then(level2).then(MetricStep())

        results = level3.run([E2EContext(sample="deep")])
        out = results[0].output
        assert out.agent_output == "answer_for_deep"
        assert out.environment_result is not None
        assert out.metadata.get("metric") is not None


# ---------------------------------------------------------------------------
# E2E: multiple run() calls on same instance
# ---------------------------------------------------------------------------


@pytest.mark.integration
class TestMultipleRunCalls:
    def test_two_run_calls_accumulate_background_work(self):
        UpdateStep._updates.clear()
        pipe = (
            Pipeline()
            .then(AgentStep())
            .then(EvaluateStep())
            .then(ReflectStep())
            .then(UpdateStep())
        )
        pipe.run([E2EContext(sample="a"), E2EContext(sample="b")])
        pipe.run([E2EContext(sample="c"), E2EContext(sample="d")])
        pipe.wait_for_background(timeout=10.0)
        assert len(UpdateStep._updates) == 4

    def test_pipeline_state_not_contaminated_between_runs(self):
        pipe = Pipeline().then(AgentStep()).then(EvaluateStep())
        r1 = pipe.run([E2EContext(sample="sample_x")])
        r2 = pipe.run([E2EContext(sample="sample_y")])
        assert r1[0].output.sample == "sample_x"
        assert r2[0].output.sample == "sample_y"


# ---------------------------------------------------------------------------
# E2E: async steps inside pipeline
# ---------------------------------------------------------------------------


@pytest.mark.integration
class TestAsyncStepsInPipeline:
    def test_async_step_runs_via_run(self):
        class AsyncAgent:
            requires = frozenset()
            provides = frozenset({"agent_output"})

            async def __call__(self, ctx: E2EContext) -> E2EContext:
                await asyncio.sleep(0)
                return ctx.replace(agent_output="async_answer")

        results = Pipeline().then(AsyncAgent()).run([E2EContext(sample="s")])
        assert results[0].output.agent_output == "async_answer"

    def test_mixed_sync_async_steps(self):
        class AsyncAgent:
            requires = frozenset()
            provides = frozenset({"agent_output"})

            async def __call__(self, ctx: E2EContext) -> E2EContext:
                await asyncio.sleep(0)
                return ctx.replace(agent_output="async")

        class SyncEval:
            requires = frozenset({"agent_output"})
            provides = frozenset({"environment_result"})

            def __call__(self, ctx: E2EContext) -> E2EContext:
                return ctx.replace(environment_result={"score": 1.0})

        results = (
            Pipeline().then(AsyncAgent()).then(SyncEval()).run([E2EContext(sample="s")])
        )
        assert results[0].output.agent_output == "async"
        assert results[0].output.environment_result["score"] == 1.0

    def test_run_async_entry_point(self):
        contexts = [E2EContext(sample="q1"), E2EContext(sample="q2")]
        results = asyncio.run(
            Pipeline().then(AgentStep()).then(EvaluateStep()).run_async(contexts)
        )
        assert len(results) == 2
        assert all(r.error is None for r in results)


# ---------------------------------------------------------------------------
# E2E: background executor shared across pipeline instances
# ---------------------------------------------------------------------------


@pytest.mark.integration
@pytest.mark.slow
class TestSharedBackgroundExecutor:
    def test_same_step_class_uses_same_executor(self):
        """Two Pipeline instances sharing the same step class share the executor."""

        class SharedBg:
            requires = frozenset()
            provides = frozenset({"done"})
            async_boundary = True
            max_workers = 1  # single-threaded shared pool
            call_count = 0
            _lock = threading.Lock()

            def __call__(self, ctx):
                with SharedBg._lock:
                    SharedBg.call_count += 1
                time.sleep(0.01)
                return ctx.replace(
                    metadata=MappingProxyType({**ctx.metadata, "done": True})
                )

        SharedBg.call_count = 0
        # Reset class executor so the test is independent
        if hasattr(SharedBg, "_executor") and SharedBg._executor is not None:
            SharedBg._executor.shutdown(wait=False)
            SharedBg._executor = None

        pipe_a = Pipeline().then(SharedBg())
        pipe_b = Pipeline().then(SharedBg())

        results_a = pipe_a.run([E2EContext(sample="a1"), E2EContext(sample="a2")])
        results_b = pipe_b.run([E2EContext(sample="b1"), E2EContext(sample="b2")])

        pipe_a.wait_for_background(timeout=5.0)
        pipe_b.wait_for_background(timeout=5.0)

        assert SharedBg.call_count == 4
        assert all(r.output.metadata.get("done") for r in results_a + results_b)


# ---------------------------------------------------------------------------
# E2E: iterable (non-list) inputs
# ---------------------------------------------------------------------------


@pytest.mark.integration
class TestIterableInputs:
    def test_generator_input(self):
        """Pipeline.run() accepts a generator, not only a list."""

        def gen():
            for s in ("a", "b", "c"):
                yield E2EContext(sample=s)

        results = Pipeline().then(AgentStep()).run(gen())
        assert len(results) == 3
        assert all(r.error is None for r in results)
        assert {r.output.agent_output for r in results} == {
            "answer_for_a",
            "answer_for_b",
            "answer_for_c",
        }

    def test_tuple_input(self):
        """Pipeline.run() accepts a tuple of contexts."""
        contexts = tuple(E2EContext(sample=s) for s in ("x", "y"))
        results = Pipeline().then(AgentStep()).then(EvaluateStep()).run(contexts)
        assert len(results) == 2
        assert all(r.error is None for r in results)

    def test_run_async_with_generator(self):
        """Pipeline.run_async() also accepts a generator."""

        def gen():
            for s in ("q1", "q2"):
                yield E2EContext(sample=s)

        results = asyncio.run(
            Pipeline().then(AgentStep()).then(EvaluateStep()).run_async(gen())
        )
        assert len(results) == 2
        assert all(r.error is None for r in results)