hermes-agent/tests/gateway/test_context_ref_expansion_runtime.py

"""Regression test for the "@" context-reference-expansion block in
``GatewayRunner._prepare_inbound_message_text``.

Bug: the block read ``self._model`` / ``self._base_url`` to resolve the
model/base_url for ``get_model_context_length_async``. ``GatewayRunner``
never assigns either attribute (that pattern was copy-pasted from
``HermesCLI``, which does carry ``self.model``/``self.base_url`` — see
commit da44c196b). Every message containing "@" raised ``AttributeError``
inside the ``try`` block, which the surrounding ``except Exception`` silently
swallowed at debug level, so ``preprocess_context_references_async`` never
ran in the gateway and @-references (``@file:``, ``@folder:``, ``@diff``,
etc.) passed through to the model completely unexpanded.

These tests pin the fix: the block must resolve model/provider/base_url via
``self._resolve_session_agent_runtime`` (the same session-aware resolution
the hygiene-compression block already uses) and must actually reach
``preprocess_context_references_async`` with a real context length.
"""
import logging
import threading
from contextlib import contextmanager

import pytest

import gateway.run as gateway_run
from agent.context_references import ContextReferenceResult
from gateway.config import GatewayConfig, Platform, PlatformConfig
from gateway.platforms.base import MessageEvent
from gateway.run import GatewayRunner
from gateway.session import SessionSource


def _make_runner() -> GatewayRunner:
    runner = object.__new__(GatewayRunner)
    runner.config = GatewayConfig(
        platforms={Platform.TELEGRAM: PlatformConfig(enabled=True, token="fake")},
    )
    runner.adapters = {}
    # Attrs touched by _resolve_session_agent_runtime on a bare test runner
    # (mirrors tests/gateway/test_empty_model_recovery.py).
    runner._session_model_overrides = {}
    runner._last_resolved_model = {}
    runner._service_tier = None
    runner._agent_cache = {}
    runner._agent_cache_lock = threading.Lock()
    return runner


def _source() -> SessionSource:
    return SessionSource(
        platform=Platform.TELEGRAM,
        chat_id="123",
        chat_name="DM",
        chat_type="private",
        user_name="Alice",
    )


def _patch_runtime_resolution(monkeypatch) -> None:
    """Stub the module-level runtime resolution so the test never hits the
    network — mirrors _patch_resolution() in test_empty_model_recovery.py."""
    monkeypatch.setattr(
        gateway_run, "_resolve_gateway_model", lambda cfg=None: "openai/gpt-4.1-mini"
    )
    monkeypatch.setattr(
        gateway_run,
        "_resolve_runtime_agent_kwargs",
        lambda: {
            "provider": "openai",
            "api_key": "test-key",
            "base_url": "https://api.openai.com/v1",
            "api_mode": "chat_completions",
        },
    )
    # config_context_length is int > 0, so get_model_context_length_async's
    # config-override short-circuit (agent/model_metadata.py) fires and
    # returns it directly — no network probe needed.
    monkeypatch.setattr(
        gateway_run,
        "_load_gateway_config",
        lambda: {"model": {"default": "openai/gpt-4.1-mini", "context_length": 128000}},
    )


@pytest.mark.asyncio
async def test_at_reference_reaches_preprocessor_with_real_context_length(
    monkeypatch, caplog
):
    """A message containing "@" must reach preprocess_context_references_async
    with a real (int > 0) context_length, and the except branch must not
    fire. This fails on unfixed code with AttributeError:
    'GatewayRunner' object has no attribute '_model' (swallowed as a debug
    log, so pre-fix this assertion sees no expansion and no captured call)."""
    runner = _make_runner()
    source = _source()
    _patch_runtime_resolution(monkeypatch)

    captured: dict = {}

    async def _fake_preprocess(message, *, cwd, context_length, url_fetcher=None, allowed_root=None):
        captured["message"] = message
        captured["cwd"] = cwd
        captured["context_length"] = context_length
        captured["allowed_root"] = allowed_root
        return ContextReferenceResult(
            message="[expanded body]",
            original_message=message,
            expanded=True,
        )

    import agent.context_references as ctx_mod

    monkeypatch.setattr(ctx_mod, "preprocess_context_references_async", _fake_preprocess)

    caplog.set_level(logging.DEBUG, logger="gateway.run")

    event = MessageEvent(text="please look at @file:notes.txt", source=source)

    result = await runner._prepare_inbound_message_text(
        event=event,
        source=source,
        history=[],
    )

    # The except branch (AttributeError on self._model/self._base_url,
    # pre-fix) must not have fired.
    assert not any(
        "@ context reference expansion failed" in record.getMessage()
        for record in caplog.records
    ), "the except branch swallowed an exception instead of reaching the preprocessor"

    # preprocess_context_references_async must actually have been called,
    # with a real positive context length (not skipped by the AttributeError).
    assert captured, "preprocess_context_references_async was never called"
    assert isinstance(captured["context_length"], int)
    assert captured["context_length"] > 0
    assert captured["context_length"] == 128000

    # The expanded result from the (stubbed) preprocessor must have been
    # adopted as the final message text.
    assert result == "[expanded body]"


@pytest.mark.asyncio
async def test_at_reference_ignores_global_context_for_runtime_route_override(monkeypatch):
    """Context expansion must not inherit a global pin from another route."""
    runner = _make_runner()
    source = _source()
    captured = {}

    monkeypatch.setattr(
        gateway_run,
        "_load_gateway_config",
        lambda: {
            "model": {
                "default": "shared-model",
                "provider": "custom",
                "base_url": "https://large.example/v1",
                "context_length": 1_048_576,
            }
        },
    )
    monkeypatch.setattr(
        runner,
        "_resolve_session_agent_runtime",
        lambda **_kwargs: (
            "shared-model",
            {
                "provider": "custom",
                "api_key": "test",
                "base_url": "https://small.example/v1",
            },
        ),
    )

    import agent.context_references as ctx_mod
    import agent.model_metadata as model_meta_mod

    async def _fake_get_context(_model, **kwargs):
        captured["config_context_length"] = kwargs["config_context_length"]
        return 32_768

    async def _passthrough(message, **_kwargs):
        return ContextReferenceResult(message=message, original_message=message)

    monkeypatch.setattr(model_meta_mod, "get_model_context_length_async", _fake_get_context)
    monkeypatch.setattr(ctx_mod, "preprocess_context_references_async", _passthrough)

    await runner._prepare_inbound_message_text(
        event=MessageEvent(text="@file:note", source=source), source=source, history=[]
    )
    assert captured["config_context_length"] is None