unsloth/studio/backend/tests/test_message_content.py

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

"""Tests for `content_to_text`, the #4383 fix for list-form message content.

Loaded by file path so the test skips importing ``core.inference`` (whose
``__init__`` pulls in the orchestrator + llama_cpp / torch).
"""

import importlib.util
from pathlib import Path


_BACKEND_DIR = Path(__file__).resolve().parent.parent


def _load_message_content():
    path = _BACKEND_DIR / "core/inference/message_content.py"
    spec = importlib.util.spec_from_file_location("message_content_under_test", path)
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)
    return module


def test_string_is_returned_unchanged():
    mc = _load_message_content()
    assert mc.content_to_text("hello world") == "hello world"
    assert mc.content_to_text("") == ""


def test_none_becomes_empty_string():
    mc = _load_message_content()
    assert mc.content_to_text(None) == ""


def test_single_text_part_list():
    mc = _load_message_content()
    content = [{"type": "text", "text": "hello"}]
    assert mc.content_to_text(content) == "hello"


def test_multimodal_list_drops_non_text_parts():
    mc = _load_message_content()
    content = [
        {"type": "text", "text": "describe this"},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
    ]
    assert mc.content_to_text(content) == "describe this"


def test_multiple_text_parts_joined_with_newline():
    mc = _load_message_content()
    content = [
        {"type": "text", "text": "first"},
        {"type": "text", "text": "second"},
    ]
    assert mc.content_to_text(content) == "first\nsecond"


def test_bare_string_items_in_list():
    mc = _load_message_content()
    assert mc.content_to_text(["a", "b"]) == "a\nb"


def test_audio_and_image_only_list_is_empty():
    mc = _load_message_content()
    content = [
        {"type": "image_url", "image_url": {"url": "x"}},
        {"type": "input_audio", "input_audio": {"data": "y", "format": "wav"}},
    ]
    assert mc.content_to_text(content) == ""


def test_part_without_type_treated_as_text():
    mc = _load_message_content()
    # A ``text`` field with no ``type`` is treated as text.
    assert mc.content_to_text([{"text": "untyped"}]) == "untyped"


def test_empty_text_parts_skipped():
    mc = _load_message_content()
    content = [
        {"type": "text", "text": ""},
        {"type": "text", "text": "kept"},
    ]
    assert mc.content_to_text(content) == "kept"


def test_tuple_behaves_like_list():
    mc = _load_message_content()
    content = ({"type": "text", "text": "x"}, {"type": "text", "text": "y"})
    assert mc.content_to_text(content) == "x\ny"


def test_result_supports_string_ops():
    mc = _load_message_content()
    # Crux of #4383: result must be a plain str for caller .strip()/.replace().
    out = mc.content_to_text([{"type": "text", "text": "  padded  "}])
    assert out.strip() == "padded"
    assert isinstance(out, str)


def _pasted(name, body, size):
    return f"<pasted_text name={name} bytes={size}>\n{body}\n</pasted_text>"


def test_paste_only_turn_is_not_empty_text():
    mc = _load_message_content()
    body = "Fix the retry backoff\ndetail"
    message = {
        "content": [],
        "attachments": [{"content": [{"type": "text", "text": _pasted("Fix.txt", body, 28)}]}],
    }
    # Deep research rejects a message whose text is empty, and a long paste
    # carries all of its text in the attachment.
    assert mc.message_text_with_pastes(message) == body


def test_typed_text_keeps_the_paste_after_it():
    mc = _load_message_content()
    message = {
        "content": [{"type": "text", "text": "summarise this"}],
        "attachments": [{"content": [{"type": "text", "text": _pasted("Log.txt", "line", 4)}]}],
    }
    assert mc.message_text_with_pastes(message) == "summarise this\n\nline"


def test_other_attachments_are_left_out():
    mc = _load_message_content()
    message = {
        "content": [{"type": "text", "text": "read this"}],
        "attachments": [
            {"content": [{"type": "text", "text": "[PDF: paper.pdf]\nAbstract"}]},
            {"content": [{"type": "text", "text": "<attachment name=n.txt>\nx\n</attachment>"}]},
        ],
    }
    assert mc.message_text_with_pastes(message) == "read this"
    assert mc.message_text_with_pastes({}) == ""
    assert mc.message_text_with_pastes(None) == ""


def test_pasted_body_unwraps_only_the_wrapper():
    mc = _load_message_content()
    assert mc.pasted_text_body(_pasted("a.txt", "body", 4)) == "body"
    # Without the size, and with a body that itself ends in a tag-like line.
    assert mc.pasted_text_body("<pasted_text name=a.txt>\nbody") == "body"
    assert mc.pasted_text_body("plain text") == ""
    assert mc.pasted_text_body("<attachment name=a.txt>\nbody\n</attachment>") == ""
    assert mc.pasted_text_body("") == ""