Skip to content
Python
23.5 KB
Raw Blame
"""P2 gate tests — turn engine + event bus (scripted provider, no network)."""

from __future__ import annotations

import asyncio
import threading
import time

import aisuite as ai
from coworker.engine import ApprovalOutcome, PermissionRequest, TurnEngine
from coworker.events import EventType
from coworker.permissions import PermissionEngine
from coworker.providers import (
    AssistantTurn,
    ModelCapabilities,
    ProviderClient,
    StreamChunk,
    ToolCall,
)
from coworker.tools import ToolRegistry


def _text_turn(text):
    return AssistantTurn(text=text, finish_reason="stop")


def _tool_turn(name, args, call_id="call_1"):
    return AssistantTurn(
        tool_calls=[ToolCall(id=call_id, name=name, arguments=args)],
        finish_reason="tool_calls",
    )


class ScriptedProvider(ProviderClient):
    """Returns queued AssistantTurns; streams via the base default (one final chunk)."""

    def __init__(self, turns, *, loop=False):
        self._turns = list(turns)
        self._loop = loop
        self.calls = 0

    def complete(self, *, model, messages, tools=None, **settings):
        self.calls += 1
        return self._turns[0] if self._loop else self._turns.pop(0)

    def capabilities(self, model):
        return ModelCapabilities()


def _engine(tmp_path, turns, *, approver=None, loop=False, max_iterations=12):
    provider = ScriptedProvider(turns, loop=loop)
    registry = ToolRegistry()
    registry.register_all(ai.toolkits.files(root=str(tmp_path), allow_write=True))
    permissions = PermissionEngine(workspace_root=tmp_path)
    engine = TurnEngine(
        provider=provider,
        registry=registry,
        permissions=permissions,
        model="gpt-5.5",
        approver=approver,
        max_iterations=max_iterations,
    )
    return engine, provider


def _collect(engine, user_input):
    async def _run():
        return [ev async for ev in engine.run(user_input)]

    return asyncio.run(_run())


def _types(events):
    return [ev.type for ev in events]


# -- tests ----------------------------------------------------------------------


def test_no_tool_turn(tmp_path):
    engine, _ = _engine(tmp_path, [_text_turn("all done")])
    events = _collect(engine, "hi")
    assert _types(events) == [
        EventType.TURN_START,
        EventType.ASSISTANT_MESSAGE,
        EventType.TURN_END,
    ]
    assert events[1].data["text"] == "all done"
    assert events[-1].data["status"] == "completed"


def test_tool_turn_order_and_execution(tmp_path):
    (tmp_path / "a.txt").write_text("hello", encoding="utf-8")
    engine, _ = _engine(
        tmp_path,
        [_tool_turn("read_file", {"path": "a.txt"}), _text_turn("it says hello")],
    )
    events = _collect(engine, "read a.txt")
    assert EventType.PERMISSION_REQUIRED not in _types(events)
    assert _types(events) == [
        EventType.TURN_START,
        EventType.ASSISTANT_MESSAGE,
        EventType.TOOL_PROPOSED,
        EventType.TOOL_STARTED,
        EventType.TOOL_FINISHED,
        EventType.ITERATION_END,
        EventType.ASSISTANT_MESSAGE,
        EventType.TURN_END,
    ]
    finished = next(e for e in events if e.type == EventType.TOOL_FINISHED)
    assert finished.data["status"] == "ok"
    assert any(
        m.get("role") == "tool" and "hello" in m["content"] for m in engine.messages
    )


def test_write_requires_approval_then_approved(tmp_path):
    async def approve_once(_req: PermissionRequest):
        return ApprovalOutcome.ONCE

    engine, _ = _engine(
        tmp_path,
        [
            _tool_turn("write_file", {"path": "new.py", "content": "print(1)\n"}),
            _text_turn("wrote new.py"),
        ],
        approver=approve_once,
    )
    events = _collect(engine, "create new.py")
    assert EventType.PERMISSION_REQUIRED in _types(events)
    assert (tmp_path / "new.py").read_text() == "print(1)\n"


def test_denied_tool_yields_error_and_continues(tmp_path):
    async def deny(_req: PermissionRequest):
        return ApprovalOutcome.DENY

    engine, _ = _engine(
        tmp_path,
        [
            _tool_turn("write_file", {"path": "new.py", "content": "x"}),
            _text_turn("ok, skipped it"),
        ],
        approver=deny,
    )
    events = _collect(engine, "create new.py")
    assert not (tmp_path / "new.py").exists()
    finished = next(e for e in events if e.type == EventType.TOOL_FINISHED)
    assert finished.data["status"] == "denied"
    assert _types(events)[-1] == EventType.TURN_END
    assert any(
        m.get("role") == "tool" and "not executed" in m["content"]
        for m in engine.messages
    )


def test_max_iterations_rail(tmp_path):
    engine, provider = _engine(
        tmp_path, [_tool_turn("list_files", {})], loop=True, max_iterations=3
    )
    events = _collect(engine, "loop forever")
    end = events[-1]
    assert end.type == EventType.TURN_END
    assert end.data["status"] == "max_iterations_exceeded"
    assert provider.calls == 3


def test_interrupt_between_iterations(tmp_path):
    engine_holder = {}

    async def approve_and_interrupt(_req: PermissionRequest):
        engine_holder["engine"].request_interrupt()
        return ApprovalOutcome.ONCE

    engine, provider = _engine(
        tmp_path,
        [
            _tool_turn("write_file", {"path": "x.py", "content": "x"}),
            _text_turn("should not be reached"),
        ],
        approver=approve_and_interrupt,
    )
    engine_holder["engine"] = engine
    events = _collect(engine, "do a thing")
    assert events[-1].type == EventType.INTERRUPTED
    assert provider.calls == 1


def test_steering_injects_next_turn(tmp_path):
    engine, provider = _engine(tmp_path, [_text_turn("first"), _text_turn("second")])
    engine.queue_steering("actually, also do this")
    events = _collect(engine, "do the first thing")
    assert provider.calls == 2
    assert any(
        m.get("role") == "user" and m["content"] == "actually, also do this"
        for m in engine.messages
    )
    assert events[-1].data["status"] == "completed"


# -- parallel tool execution ------------------------------------------------------


def _multi_tool_turn(calls):
    return AssistantTurn(
        tool_calls=[
            ToolCall(id=f"call_{i}", name=name, arguments=args)
            for i, (name, args) in enumerate(calls)
        ],
        finish_reason="tool_calls",
    )


def _bare_engine(tmp_path, turns):
    provider = ScriptedProvider(turns)
    registry = ToolRegistry()
    permissions = PermissionEngine(workspace_root=tmp_path)
    engine = TurnEngine(
        provider=provider,
        registry=registry,
        permissions=permissions,
        model="gpt-5.5",
    )
    return engine, registry


def test_low_risk_tool_calls_run_concurrently(tmp_path):
    # Both tools block on a 2-party barrier: the turn only completes if the engine
    # really runs them at the same time (sequential execution would trip the timeout
    # and surface as an error result).
    barrier = threading.Barrier(2, timeout=5)
    low = ai.ToolMetadata(category="search", risk_level="low", requires_approval=False)

    def side_a():
        """Wait for side_b."""
        barrier.wait()
        return {"side": "a"}

    def side_b():
        """Wait for side_a."""
        barrier.wait()
        return {"side": "b"}

    engine, registry = _bare_engine(
        tmp_path,
        [_multi_tool_turn([("side_a", {}), ("side_b", {})]), _text_turn("done")],
    )
    registry.register(side_a, metadata=low)
    registry.register(side_b, metadata=low)

    events = _collect(engine, "go")
    finished = [e for e in events if e.type == EventType.TOOL_FINISHED]
    assert len(finished) == 2
    assert all(e.data["status"] == "ok" for e in finished)
    # a tool result message exists for every call id
    tool_ids = {
        m.get("tool_call_id") for m in engine.messages if m.get("role") == "tool"
    }
    assert tool_ids == {"call_0", "call_1"}


def test_non_low_risk_tool_calls_stay_sequential(tmp_path):
    order = []
    medium = ai.ToolMetadata(
        category="filesystem", risk_level="medium", requires_approval=False
    )

    def first():
        """Record start/end with a delay."""
        order.append("first-start")
        time.sleep(0.2)
        order.append("first-end")
        return "ok"

    def second():
        """Record start/end."""
        order.append("second-start")
        order.append("second-end")
        return "ok"

    engine, registry = _bare_engine(
        tmp_path,
        [_multi_tool_turn([("first", {}), ("second", {})]), _text_turn("done")],
    )
    registry.register(first, metadata=medium)
    registry.register(second, metadata=medium)

    _collect(engine, "go")
    assert order == ["first-start", "first-end", "second-start", "second-end"]


class StreamingProvider(ProviderClient):
    def complete(self, **kwargs):  # pragma: no cover - streamed instead
        raise NotImplementedError

    def capabilities(self, model):
        return ModelCapabilities()

    def stream(self, *, model, messages, tools=None, **settings):
        for piece in ["Hel", "lo, ", "world"]:
            yield StreamChunk(text_delta=piece)
        yield StreamChunk(turn=AssistantTurn(text="Hello, world", finish_reason="stop"))


def test_streaming_emits_deltas(tmp_path):
    registry = ToolRegistry()
    permissions = PermissionEngine(workspace_root=tmp_path)
    engine = TurnEngine(
        provider=StreamingProvider(),
        registry=registry,
        permissions=permissions,
        model="gpt-5.5",
    )
    events = _collect(engine, "say hi")
    deltas = [e.data["text"] for e in events if e.type == EventType.ASSISTANT_DELTA]
    assert deltas == ["Hel", "lo, ", "world"]
    final = next(e for e in events if e.type == EventType.ASSISTANT_MESSAGE)
    assert final.data["text"] == "Hello, world"
    assert events[-1].type == EventType.TURN_END


class DyingStreamProvider(ProviderClient):
    """Streams a delta, then dies — the shape of a provider 503 landing mid-turn."""

    def __init__(self):
        self.payloads = []

    def complete(self, **kwargs):  # pragma: no cover - streamed instead
        raise NotImplementedError

    def capabilities(self, model):
        return ModelCapabilities()

    def stream(self, *, model, messages, tools=None, **settings):
        self.payloads.append(messages)
        yield StreamChunk(text_delta="I will request permission to ")
        raise RuntimeError("Error code: 503 - {'error': {'type': 'overloaded_error'}}")


def test_retry_after_mid_stream_death_does_not_replay_the_partial(tmp_path):
    # The partial stays in the transcript so the user can see what streamed, but several
    # providers reject a payload whose FINAL message is an assistant turn — Gemini ("Requests
    # ending with a model turn are not supported"), Claude 4.6+ ("does not support assistant
    # message prefill"), older Claude when it ends in whitespace. Replaying it wedges retry.
    provider = DyingStreamProvider()
    engine = TurnEngine(
        provider=provider,
        registry=ToolRegistry(),
        permissions=PermissionEngine(workspace_root=tmp_path),
        model="gpt-5.5",
    )
    _collect(engine, "arrange my visa documents")

    # kept in history (display), followed by the error notice
    assert engine.messages[-2]["role"] == "assistant"
    assert engine.messages[-2]["content"] == "I will request permission to "
    assert engine.messages[-1]["role"] == "notice"

    async def _retry():
        return [ev async for ev in engine.retry()]

    asyncio.run(_retry())

    sent = provider.payloads[-1]
    assert sent, "retry never reached the provider"
    assert sent[-1]["role"] != "assistant"
    assert all(m.get("content") != "I will request permission to " for m in sent)


def test_completed_assistant_turn_is_never_dropped(tmp_path):
    # The counterpart to the test above: a finished turn's assistant message is structurally
    # identical to an abandoned partial (trailing, no tool_calls). Only the error notice tells
    # them apart, so a completed turn with no notice after it must survive intact.
    engine, _ = _engine(tmp_path, [_text_turn("all done")])
    _collect(engine, "hi")

    assert engine.messages[-1]["role"] == "assistant"
    outbound = engine._outbound_messages()
    assert outbound[-1]["role"] == "assistant"
    assert outbound[-1]["content"] == "all done"


def test_consecutive_deaths_do_not_stack_partials_in_the_payload(tmp_path):
    # Each failed retry appends its own partial, so one drop is not enough.
    provider = DyingStreamProvider()
    engine = TurnEngine(
        provider=provider,
        registry=ToolRegistry(),
        permissions=PermissionEngine(workspace_root=tmp_path),
        model="gpt-5.5",
    )
    _collect(engine, "go")

    async def _retry():
        return [ev async for ev in engine.retry()]

    asyncio.run(_retry())
    asyncio.run(_retry())

    assert sum(1 for m in engine.messages if m.get("role") == "assistant") == 3
    assert provider.payloads[-1][-1]["role"] != "assistant"


def _pdf_file_part():
    import base64
    import io

    from pypdf import PdfWriter

    writer = PdfWriter()
    writer.add_blank_page(width=100, height=100)
    buf = io.BytesIO()
    writer.write(buf)
    url = "data:application/pdf;base64," + base64.b64encode(buf.getvalue()).decode()
    return {"type": "file", "file": {"filename": "d.pdf", "file_data": url}}


def test_outbound_adapts_pdf_for_non_pdf_models(tmp_path):
    # ScriptedProvider reports default caps (pdf=False) → the file part must be
    # replaced at send time while the stored history keeps the real document.
    engine, _ = _engine(tmp_path, [_text_turn("ok")])
    engine.messages.append(
        {
            "role": "user",
            "content": [{"type": "text", "text": "read this"}, _pdf_file_part()],
        }
    )
    parts = engine._outbound_messages()[-1]["content"]
    assert all(p["type"] != "file" for p in parts)
    assert "d.pdf" in parts[-1]["text"]
    assert engine.messages[-1]["content"][1]["type"] == "file"  # history untouched


def test_outbound_keeps_pdf_for_native_models(tmp_path):
    class NativeProvider(ScriptedProvider):
        def capabilities(self, model):
            return ModelCapabilities(vision=True, pdf=True)

    engine, _ = _engine(tmp_path, [_text_turn("ok")])
    engine.provider = NativeProvider([_text_turn("ok")])
    message = {
        "role": "user",
        "content": [{"type": "text", "text": "read this"}, _pdf_file_part()],
    }
    engine.messages.append(message)
    assert engine._outbound_messages()[-1]["content"][1]["type"] == "file"


def test_provider_extras_persist_on_message_and_survive_outbound(tmp_path):
    """A turn's provider-private sidecar (`extras`, e.g. Gemini thought signatures) rides
    the persisted assistant message and is NOT stripped by _outbound_messages — the owning
    provider needs it back; foreign providers strip it themselves."""
    turn = AssistantTurn(
        text="ok",
        finish_reason="stop",
        extras={"_gemini": {"text_sig": "c2ln", "call_sigs": []}},
    )
    engine, _ = _engine(tmp_path, [turn])
    _collect(engine, "hi")

    persisted = engine.messages[-1]
    assert persisted["_gemini"] == {"text_sig": "c2ln", "call_sigs": []}
    outbound = engine._outbound_messages()[-1]
    assert outbound["_gemini"] == {"text_sig": "c2ln", "call_sigs": []}
    assert "ts" not in outbound  # display sidecars still stripped


def test_switch_model_appends_notice_only_midsession(tmp_path):
    engine, _ = _engine(tmp_path, [_text_turn("ok")])
    # Fresh session: first bind is silent.
    assert engine.switch_model("zai:glm-5.2") is None
    assert engine.model == "zai:glm-5.2"
    _collect(engine, "hi")
    # Same model: no-op.
    assert engine.switch_model("zai:glm-5.2") is None
    # Real mid-session switch: persisted marker with the matrix label.
    text = engine.switch_model("kimi:kimi-k2.6")
    assert "Kimi K2.6" in text and engine.model == "kimi:kimi-k2.6"
    notice = engine.messages[-1]
    assert notice["role"] == "notice" and notice["kind"] == "model_switch"
    assert all(m.get("role") != "notice" for m in engine._outbound_messages())


def test_switch_model_warns_when_images_meet_text_only_model(tmp_path):
    class NoVisionProvider(ScriptedProvider):
        def capabilities(self, model):
            return ModelCapabilities(vision=False)

    engine, _ = _engine(tmp_path, [_text_turn("ok")])
    engine.provider = NoVisionProvider([_text_turn("ok")])
    engine.messages.append(
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "look"},
                {"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}},
            ],
        }
    )
    text = engine.switch_model("zai:glm-5.2")
    assert "images" in text  # degradation is called out in the marker


def test_outbound_replaces_images_for_non_vision_models(tmp_path):
    class NoVisionProvider(ScriptedProvider):
        def capabilities(self, model):
            return ModelCapabilities(vision=False)

    engine, _ = _engine(tmp_path, [_text_turn("ok")])
    engine.provider = NoVisionProvider([_text_turn("ok")])
    engine.messages.append(
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "look"},
                {"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}},
            ],
        }
    )
    parts = engine._outbound_messages()[-1]["content"]
    assert all(p["type"] != "image_url" for p in parts)
    assert "not viewable" in parts[-1]["text"]
    assert engine.messages[-1]["content"][1]["type"] == "image_url"  # history untouched


def test_new_turn_repairs_dangling_tool_calls(tmp_path):
    """A restart mid-approval leaves an assistant tool_use with no result;
    providers 400 the whole conversation for it, poisoning the session (found
    live in the machine-events drill). A NEW user turn stubs every orphan —
    in place, id-matched, honest — before appending its message."""
    engine, _ = _engine(tmp_path, [_text_turn("recovered")])
    engine.messages.append(
        {
            "role": "assistant",
            "content": "",
            "tool_calls": [
                {
                    "id": "call_lost",
                    "type": "function",
                    "function": {"name": "send_message", "arguments": "{}"},
                }
            ],
        }
    )
    events = _collect(engine, "follow-up")
    assert events[-1].data["status"] == "completed"
    idx = next(
        i
        for i, m in enumerate(engine.messages)
        if m.get("role") == "assistant" and m.get("tool_calls")
    )
    stub = engine.messages[idx + 1]
    assert stub["role"] == "tool" and stub["tool_call_id"] == "call_lost"
    assert "interrupted" in stub["content"]
    # The user's follow-up comes AFTER the stub — valid provider ordering.
    assert engine.messages[idx + 2]["role"] == "user"


def test_leaked_tool_call_ends_the_turn_as_a_retriable_error(tmp_path):
    """A tool call the endpoint couldn't parse must not pass as an answer. Ending "completed"
    made a half-written call indistinguishable from the model deciding it was done — the user
    saw narration trailing off into stray tags (owner report 2026-07-26, qwen3.5-9b on LM
    Studio). It ends on the error path so the GUI offers Retry; the drift is probabilistic, so
    retrying the same model usually works."""
    leaked = "Let me read the key files.\n<tool_call>\n<function=nope_not_a_tool>\n<parameter="
    engine, _ = _engine(tmp_path, [_text_turn(leaked)])
    events = _collect(engine, "explore the codebase")

    assert EventType.ERROR in _types(events)
    assert EventType.TURN_END not in _types(events)
    err = next(ev for ev in events if ev.type == EventType.ERROR)
    assert err.data["error_type"] == "UnparsedToolCall"
    assert "couldn't parse" in err.data["error"]
    # Persisted as an error notice, which is what unlocks retry().
    assert engine.messages[-1] == {
        **engine.messages[-1],
        "role": "notice",
        "kind": "error",
    }
    assert engine._tail_is_retriable_error() is True


def test_ordinary_text_answer_still_completes(tmp_path):
    """Guard the other side: prose that merely mentions tool syntax inside code fences is a
    real answer and must still complete normally."""
    engine, _ = _engine(
        tmp_path,
        [_text_turn("Qwen writes calls like:\n```\n<tool_call><function=x>\n```\nThat's it.")],
    )
    events = _collect(engine, "how does qwen format tool calls?")
    assert EventType.ERROR not in _types(events)
    assert next(ev for ev in events if ev.type == EventType.TURN_END).data["status"] == "completed"


# -- audit correlation -------------------------------------------------------------


def test_every_audit_row_carries_the_tool_call_id(tmp_path):
    """#670: an external audit sink can only join a `started` to its `finished` if each
    row names the call it belongs to. The ID is stamped at the shared `_audit` boundary,
    so this covers the lifecycle stages that never pass it themselves."""
    (tmp_path / "a.txt").write_text("hello", encoding="utf-8")
    rows: list[dict] = []
    provider = ScriptedProvider(
        [_tool_turn("read_file", {"path": "a.txt"}, call_id="call_a"), _text_turn("done")]
    )
    registry = ToolRegistry()
    registry.register_all(ai.toolkits.files(root=str(tmp_path), allow_write=True))
    engine = TurnEngine(
        provider=provider,
        registry=registry,
        permissions=PermissionEngine(workspace_root=tmp_path),
        model="gpt-5.5",
        audit_sink=rows.append,
    )
    _collect(engine, "read a.txt")

    tool_rows = [r for r in rows if r.get("stage") in {"proposed", "started", "finished"}]
    assert tool_rows, rows
    assert {r["call_id"] for r in tool_rows} == {"call_a"}
    # The `started`/`finished` pair for this call is unambiguous even when the same
    # tool and arguments run twice: each row names which invocation it is.
    assert [r["stage"] for r in tool_rows] == ["proposed", "started", "finished"]


def test_repeated_identical_calls_stay_distinguishable(tmp_path):
    """The regression that motivated #670: the same tool with the same arguments invoked
    twice in one turn produces audit rows a sink can tell apart only by call_id."""
    (tmp_path / "a.txt").write_text("hello", encoding="utf-8")
    rows: list[dict] = []
    provider = ScriptedProvider(
        [
            _tool_turn("read_file", {"path": "a.txt"}, call_id="call_1"),
            _tool_turn("read_file", {"path": "a.txt"}, call_id="call_2"),
            _text_turn("done"),
        ]
    )
    registry = ToolRegistry()
    registry.register_all(ai.toolkits.files(root=str(tmp_path), allow_write=True))
    engine = TurnEngine(
        provider=provider,
        registry=registry,
        permissions=PermissionEngine(workspace_root=tmp_path),
        model="gpt-5.5",
        audit_sink=rows.append,
    )
    _collect(engine, "read a.txt twice")

    finished = [r for r in rows if r.get("stage") == "finished"]
    assert len(finished) == 2
    assert all(r["status"] == "ok" for r in finished), finished
    assert [r["call_id"] for r in finished] == ["call_1", "call_2"]


def test_audit_call_id_is_overridable_by_a_call_site(tmp_path):
    """`**event` still wins: a call site that means something else by `call_id` is not
    clobbered by the default stamp."""
    engine, _ = _engine(tmp_path, [_text_turn("hi")])
    seen: list[dict] = []
    engine.audit_sink = seen.append
    engine._audit(ToolCall(id="real_call", name="read_file", arguments={}), stage="proposed")
    assert seen[0]["call_id"] == "real_call"
    engine._audit(
        ToolCall(id="real_call", name="read_file", arguments={}),
        stage="custom",
        call_id="other_id",
    )
    assert seen[1]["call_id"] == "other_id"