omnigent-ai--omnigent/tests/runtime/test_content_resolver.py

"""Tests for omnigent.runtime.content_resolver."""

from __future__ import annotations

import base64
from dataclasses import dataclass
from typing import Any

import pytest

from omnigent.entities import ConversationItem, StoredFile
from omnigent.entities.conversation import (
    FunctionCallData,
    MessageData,
)
from omnigent.runtime.content_resolver import resolve_content_references

# ── Fake stores ──────────────────────────────────────────────────────


@dataclass
class FakeFileStore:
    """
    In-memory FileStore that returns pre-configured StoredFile
    objects by file_id.

    :param files: Mapping of file_id to StoredFile.
    """

    files: dict[str, StoredFile]

    def get(self, file_id: str) -> StoredFile | None:
        """
        Return the StoredFile for *file_id*, or ``None``.

        :param file_id: The file identifier to look up.
        :returns: The matching StoredFile or None.
        """
        return self.files.get(file_id)


@dataclass
class FakeArtifactStore:
    """
    In-memory ArtifactStore that returns pre-configured binary
    blobs by key.

    :param blobs: Mapping of artifact key to binary content.
    """

    blobs: dict[str, bytes]

    def get(self, key: str) -> bytes:
        """
        Return binary content for *key*.

        :param key: The artifact key to look up.
        :returns: The binary content.
        :raises KeyError: If no blob exists for the key.
        """
        return self.blobs[key]


# ── Helpers ──────────────────────────────────────────────────────────


def _make_conversation_item(
    content: list[dict[str, Any]],
    *,
    role: str = "user",
    item_id: str = "msg_001",
    response_id: str = "resp_001",
) -> ConversationItem:
    """
    Build a message ConversationItem with the given content blocks.

    :param content: Content block dicts, e.g.
        ``[{"type": "input_text", "text": "Hello"}]``.
    :param role: Message role, ``"user"`` or ``"assistant"``.
    :param item_id: Store-assigned item ID.
    :param response_id: The response/task ID.
    :returns: A ConversationItem with type ``"message"``.
    """
    data = MessageData(
        role=role,
        content=content,
        # assistant messages require an agent name.
        agent="test-agent" if role == "assistant" else None,
    )
    return ConversationItem(
        id=item_id,
        type="message",
        status="completed",
        response_id=response_id,
        created_at=1000,
        data=data,
    )


def _make_function_call_item(
    *,
    item_id: str = "fc_001",
    response_id: str = "resp_001",
) -> ConversationItem:
    """
    Build a function_call ConversationItem.

    :param item_id: Store-assigned item ID.
    :param response_id: The response/task ID.
    :returns: A ConversationItem with type ``"function_call"``.
    """
    return ConversationItem(
        id=item_id,
        type="function_call",
        status="completed",
        response_id=response_id,
        created_at=1000,
        data=FunctionCallData(
            agent="test-agent",
            call_id="call_001",
            name="grep",
            arguments="{}",
        ),
    )


PNG_BYTES = b"\x89PNG\r\n\x1a\n fake png content"
PDF_BYTES = b"%PDF-1.4 fake pdf content"


@pytest.fixture()
def file_store() -> FakeFileStore:
    """
    FileStore with two pre-configured files: an image and a PDF.

    :returns: A FakeFileStore with ``file_img`` and ``file_pdf``.
    """
    return FakeFileStore(
        files={
            "file_img": StoredFile(
                id="file_img",
                created_at=1000,
                filename="photo.png",
                bytes=len(PNG_BYTES),
                content_type="image/png",
            ),
            "file_pdf": StoredFile(
                id="file_pdf",
                created_at=1000,
                filename="report.pdf",
                bytes=len(PDF_BYTES),
                content_type="application/pdf",
            ),
        }
    )


@pytest.fixture()
def artifact_store() -> FakeArtifactStore:
    """
    ArtifactStore with binary content for the image and PDF files.

    :returns: A FakeArtifactStore with blobs for ``file_img`` and
        ``file_pdf``.
    """
    return FakeArtifactStore(
        blobs={
            "file_img": PNG_BYTES,
            "file_pdf": PDF_BYTES,
        }
    )


# ── Tests ────────────────────────────────────────────────────────────


def test_text_only_message_passes_through(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    Messages with only text blocks should pass through unchanged
    (no copy, no modification).
    """
    item = _make_conversation_item([{"type": "input_text", "text": "Hello"}])
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # Exactly one item returned, same object (no copy needed).
    # Failure would mean text-only messages are unnecessarily copied.
    assert len(result) == 1
    assert result[0] is item


def test_input_image_file_id_resolved_to_data_uri(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    input_image with file_id must be resolved to a data: URI in
    image_url, with file_id removed from the block.
    """
    item = _make_conversation_item(
        [
            {"type": "input_text", "text": "What's in this image?"},
            {
                "type": "input_image",
                "file_id": "file_img",
                "detail": "auto",
            },
        ]
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # Result is a copy — original must not be modified.
    # Failure would mean in-place mutation corrupts conversation store.
    assert result[0] is not item
    assert isinstance(result[0].data, MessageData)
    blocks = result[0].data.content

    # Text block passes through unchanged.
    assert blocks[0] == {"type": "input_text", "text": "What's in this image?"}

    # Image block: file_id replaced with data: URI.
    img_block = blocks[1]
    expected_b64 = base64.b64encode(PNG_BYTES).decode("ascii")
    # file_id must be removed — it's a local reference the LLM can't use.
    # Failure would mean the LLM receives a meaningless file_id.
    assert "file_id" not in img_block
    # image_url must be a data: URI with the correct content type.
    # Failure would mean the LLM receives an invalid image reference.
    assert img_block["image_url"] == (f"data:image/png;base64,{expected_b64}")
    # detail field must be preserved — it controls provider image resolution.
    # Failure would mean the client's detail preference is lost.
    assert img_block["detail"] == "auto"
    # Block type must be preserved for downstream translation layers.
    assert img_block["type"] == "input_image"


def test_input_file_file_id_resolved_to_file_data(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    input_file with file_id must be resolved to file_data (base64),
    with content_type from file store metadata.
    """
    item = _make_conversation_item(
        [
            {"type": "input_text", "text": "Summarize"},
            {
                "type": "input_file",
                "file_id": "file_pdf",
                "filename": "report.pdf",
            },
        ]
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    assert result[0] is not item
    assert isinstance(result[0].data, MessageData)
    file_block = result[0].data.content[1]

    expected_b64 = base64.b64encode(PDF_BYTES).decode("ascii")
    # file_id must be removed.
    assert "file_id" not in file_block
    # file_data must be a data: URI with the content_type and base64.
    # Failure would mean the LLM receives no file content.
    assert file_block["file_data"] == f"data:application/pdf;base64,{expected_b64}"
    # content_type is embedded in the data: URI, not a separate field.
    assert "content_type" not in file_block
    # filename must be preserved from the original block.
    assert file_block["filename"] == "report.pdf"
    assert file_block["type"] == "input_file"


def test_image_url_passes_through_unchanged(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    input_image with image_url (no file_id) must pass through
    unchanged — URLs are never fetched server-side (SSRF protection).
    """
    item = _make_conversation_item(
        [
            {
                "type": "input_image",
                "image_url": "https://example.com/photo.png",
                "detail": "high",
            },
        ]
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # No file_id in any block — original item returned as-is.
    # Failure would mean URL-only messages are unnecessarily copied.
    assert result[0] is item


def test_inline_file_data_passes_through_unchanged(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    input_file with file_data (no file_id) must pass through
    unchanged — content is already inline.
    """
    item = _make_conversation_item(
        [
            {
                "type": "input_file",
                "file_data": "JVBERi0xLjQK",
                "filename": "report.pdf",
            },
        ]
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # No file_id — original item returned.
    assert result[0] is item


def test_non_message_items_pass_through(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    Non-message items (function_call, function_call_output) must
    pass through unchanged — only message content blocks are scanned.
    """
    fc_item = _make_function_call_item()
    result = resolve_content_references(
        [fc_item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # function_call items have no content blocks to resolve.
    # Failure would mean non-message items are incorrectly processed.
    assert len(result) == 1
    assert result[0] is fc_item


def test_missing_file_id_raises_value_error(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    Referencing a file_id that doesn't exist in the file store
    must raise ValueError — fail loud, no silent dropping.
    """
    item = _make_conversation_item(
        [
            {
                "type": "input_image",
                "file_id": "file_nonexistent",
            },
        ]
    )

    with pytest.raises(ValueError, match="file_nonexistent"):
        resolve_content_references(
            [item],
            file_store,
            artifact_store,  # type: ignore[arg-type]
        )


def test_original_item_not_mutated(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    The original ConversationItem must not be mutated — the resolver
    returns copies for modified items.
    """
    original_content = [
        {"type": "input_image", "file_id": "file_img"},
    ]
    item = _make_conversation_item(original_content)

    resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # Original content must still contain file_id — not mutated.
    # Failure would mean the resolver modifies conversation store data.
    assert isinstance(item.data, MessageData)
    assert item.data.content[0]["file_id"] == "file_img"
    assert "image_url" not in item.data.content[0]


def test_unknown_block_type_with_file_id_resolved(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    An unrecognized block type (e.g. input_audio) with file_id must
    still have its file_id resolved — the resolver resolves file_id
    on any block type, not just known ones.
    """
    item = _make_conversation_item(
        [
            {
                "type": "input_audio",
                "file_id": "file_img",
            },
        ]
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    assert isinstance(result[0].data, MessageData)
    block = result[0].data.content[0]
    # file_id resolved even for unknown type.
    # Failure would mean new content types can't use file_id.
    assert "file_id" not in block
    # Unknown types get file_data (not image_url, which is
    # only for input_image).
    assert "file_data" in block
    assert block["type"] == "input_audio"


def test_mixed_items_preserves_order(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    A mixed list of message and non-message items must preserve
    chronological order after resolution.
    """
    items = [
        _make_conversation_item(
            [{"type": "input_text", "text": "first"}],
            item_id="msg_001",
        ),
        _make_conversation_item(
            [{"type": "input_image", "file_id": "file_img"}],
            item_id="msg_002",
        ),
        _make_function_call_item(item_id="fc_001"),
        _make_conversation_item(
            [{"type": "input_text", "text": "third"}],
            item_id="msg_003",
        ),
    ]
    result = resolve_content_references(
        items,
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # Order must be preserved: msg, msg(resolved), fc, msg.
    # Failure would mean resolution reorders conversation history.
    assert len(result) == 4
    assert result[0].id == "msg_001"
    assert result[1].id == "msg_002"
    assert result[2].id == "fc_001"
    assert result[3].id == "msg_003"

    # Only msg_002 should be a copy (it had file_id).
    assert result[0] is items[0]
    assert result[1] is not items[1]
    assert result[2] is items[2]
    assert result[3] is items[3]


@pytest.mark.parametrize(
    ("block_type", "expected_field"),
    [
        pytest.param(
            "input_image",
            "image_url",
            id="input_image_gets_data_uri",
        ),
        pytest.param(
            "input_file",
            "file_data",
            id="input_file_gets_file_data",
        ),
    ],
)
def test_resolution_field_varies_by_block_type(
    block_type: str,
    expected_field: str,
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    input_image blocks get image_url (data: URI), while input_file
    blocks get file_data (data: URI). The resolution target field
    depends on block type.
    """
    item = _make_conversation_item(
        [
            {"type": block_type, "file_id": "file_img"},
        ]
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    assert isinstance(result[0].data, MessageData)
    block = result[0].data.content[0]
    # The expected field must be present after resolution.
    # Failure would mean the wrong inline format is used for this type.
    assert expected_field in block
    assert "file_id" not in block


# ── Cache tests ─────────────────────────────────────────────────────


def test_cache_avoids_redundant_artifact_fetch(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    When a cache dict is provided, the second resolution of the same
    file_id must use the cached base64 instead of re-fetching from
    the artifact store.
    """
    cache: dict[str, str] = {}
    item = _make_conversation_item([{"type": "input_image", "file_id": "file_img"}])

    # First call — populates cache.
    resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
        cache,
    )

    # Cache should now contain the file_id.
    assert "file_img" in cache
    expected_b64 = base64.b64encode(PNG_BYTES).decode("ascii")
    # Cached value is the raw base64, not a data: URI.
    assert cache["file_img"] == expected_b64

    # Sabotage the artifact store — if the cache is working, the
    # resolver won't call artifact_store.get() again.
    artifact_store.blobs = {}

    result2 = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
        cache,
    )

    # Second call should still succeed using cached value.
    assert isinstance(result2[0].data, MessageData)
    img_block = result2[0].data.content[0]
    assert img_block["image_url"] == f"data:image/png;base64,{expected_b64}"


def test_cache_none_disables_caching(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    Passing ``cache=None`` (the default) must still resolve
    correctly — caching is optional.
    """
    item = _make_conversation_item([{"type": "input_image", "file_id": "file_img"}])

    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
        # Explicit None — no cache.
        None,
    )

    assert isinstance(result[0].data, MessageData)
    img_block = result[0].data.content[0]
    expected_b64 = base64.b64encode(PNG_BYTES).decode("ascii")
    assert img_block["image_url"] == f"data:image/png;base64,{expected_b64}"


# ── Error handling tests ────────────────────────────────────────────


def test_assistant_message_file_id_resolved(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    An assistant message in conversation history with file_id must
    have that reference resolved — the content resolver scans all
    message roles, not just user messages.

    Catches bugs where role-based filtering causes assistant messages
    (e.g. from a prior turn's multimodal output) to retain
    unresolved file_id references.
    """
    item = _make_conversation_item(
        [
            {"type": "input_text", "text": "Here is the analysis"},
            {"type": "input_file", "file_id": "file_pdf", "filename": "report.pdf"},
        ],
        role="assistant",
    )
    result = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )

    # Must be a copy — assistant message was modified.
    assert result[0] is not item
    assert isinstance(result[0].data, MessageData)
    file_block = result[0].data.content[1]

    expected_b64 = base64.b64encode(PDF_BYTES).decode("ascii")
    # file_id must be removed.
    assert "file_id" not in file_block
    # file_data must be a data: URI.
    assert file_block["file_data"] == f"data:application/pdf;base64,{expected_b64}"
    assert file_block["filename"] == "report.pdf"
    assert file_block["type"] == "input_file"


def test_deleted_file_raises_clear_error(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    When a file is deleted between request validation and agent loop
    execution, the error message must clearly indicate the file was
    deleted — not a generic "not found".
    """
    item = _make_conversation_item([{"type": "input_image", "file_id": "file_nonexistent"}])

    with pytest.raises(ValueError, match="no longer exists"):
        resolve_content_references(
            [item],
            file_store,
            artifact_store,  # type: ignore[arg-type]
        )


# ── _resolve_content_type tests ───────────────────────────────────────


def test_resolve_content_type_uses_stored_type() -> None:
    """Stored content_type is used when it's not the generic fallback."""
    from omnigent.runtime.content_resolver import _resolve_content_type

    assert _resolve_content_type("text/markdown", "test.md") == "text/markdown"
    assert _resolve_content_type("application/pdf", "report.pdf") == "application/pdf"


def test_resolve_content_type_ignores_octet_stream() -> None:
    """application/octet-stream is treated as unresolved — falls through to filename."""
    from omnigent.runtime.content_resolver import _resolve_content_type

    result = _resolve_content_type("application/octet-stream", "readme.md")
    assert result == "text/markdown", f"Expected text/markdown, got {result}"


def test_resolve_content_type_falls_back_to_extra_map() -> None:
    """Extensions missing from mimetypes use the _EXTRA_MIME_TYPES fallback."""
    from omnigent.runtime.content_resolver import _resolve_content_type

    # These extensions are NOT in Python 3.10's mimetypes module
    # (except .ts, which mimetypes maps to video/mp2t).
    assert _resolve_content_type(None, "config.yaml") == "text/yaml"
    assert _resolve_content_type(None, "config.yml") == "text/yaml"
    assert _resolve_content_type(None, "main.go") == "text/x-go"
    # .ts is mapped by mimetypes to video/mp2t (MPEG transport stream)
    # and .rs to application/rls-services+xml — both wrong for source
    # code. Our _EXTRA_MIME_TYPES map takes priority.
    assert _resolve_content_type(None, "app.ts") == "text/typescript"
    assert _resolve_content_type(None, "app.tsx") == "text/typescript"
    assert _resolve_content_type(None, "lib.rs") == "text/x-rust"
    assert _resolve_content_type(None, "main.swift") == "text/x-swift"
    assert _resolve_content_type(None, "data.jsonl") == "application/jsonl"
    assert _resolve_content_type(None, "notebook.ipynb") == "application/x-ipynb+json"


def test_resolve_content_type_no_filename_uses_fallback() -> None:
    """When neither stored type nor filename is available, falls back to octet-stream."""
    from omnigent.runtime.content_resolver import _resolve_content_type

    assert _resolve_content_type(None, None) == "application/octet-stream"


def test_resolve_content_type_known_extension_uses_mimetypes() -> None:
    """Standard extensions (e.g. .pdf, .html) use mimetypes.guess_type."""
    from omnigent.runtime.content_resolver import _resolve_content_type

    # .pdf and .html are universally known by mimetypes.
    result = _resolve_content_type(None, "report.pdf")
    assert result == "application/pdf"
    result = _resolve_content_type(None, "page.html")
    assert result == "text/html"


def test_resolve_content_type_case_insensitive() -> None:
    """Extension matching is case-insensitive."""
    from omnigent.runtime.content_resolver import _resolve_content_type

    assert _resolve_content_type(None, "README.MD") == "text/markdown"
    assert _resolve_content_type(None, "config.YAML") == "text/yaml"


def test_content_resolver_uses_filename_for_mime(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """
    End-to-end: when stored content_type is octet-stream, the resolver
    uses the filename to determine the correct MIME type in the data URI.
    """
    md_content = b"# Hello\n\nMarkdown file."
    file_store.files["file_md"] = StoredFile(
        id="file_md",
        filename="readme.md",
        bytes=len(md_content),
        content_type="application/octet-stream",
        created_at=1000,
    )
    artifact_store.blobs["file_md"] = md_content

    item = _make_conversation_item(
        [{"type": "input_file", "file_id": "file_md", "filename": "readme.md"}]
    )
    resolved = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )
    resolved_data = resolved[0].data
    assert isinstance(resolved_data, MessageData)
    file_block = resolved_data.content[0]
    # Must be text/markdown, NOT application/octet-stream.
    assert file_block["file_data"].startswith("data:text/markdown;base64,"), (
        f"Expected text/markdown data URI, got: {file_block['file_data'][:60]}"
    )


# ── _safe_file_data_mime + end-to-end downgrade behaviour ────────────


@pytest.mark.parametrize(
    "passthrough_mime",
    [
        "application/pdf",
        "text/plain",
        "text/markdown",
        "text/html",
        "text/csv",
        "application/json",
        "text/javascript",
        "application/javascript",
        "text/x-python",
    ],
)
def test_safe_file_data_mime_passthrough(passthrough_mime: str) -> None:
    """MIMEs on the passthrough list are returned unchanged.

    Pinned to the explicit list because adding/removing entries changes
    what the OpenAI Responses API will accept on file_data; that's a
    deliberate decision, not an accidental one.
    """
    from omnigent.runtime.content_resolver import _safe_file_data_mime

    assert _safe_file_data_mime(passthrough_mime) == passthrough_mime


@pytest.mark.parametrize(
    "rejected_mime",
    [
        # The exact MIME the user reported (text/yaml triggers a 400
        # "unsupported MIME type 'text/yaml'" from the Responses API).
        "text/yaml",
        # Other text/* entries from _EXTRA_MIME_TYPES that the
        # Responses API rejects in the same way.
        "text/typescript",
        "text/x-rust",
        "text/x-go",
        "text/x-swift",
        "text/x-kotlin",
        "text/x-haskell",
        "text/x-julia",
        # JSONL / ndjson / ipynb — application/x-* variants the
        # Responses API treats the same way as exotic text MIMEs.
        "application/jsonl",
        "application/x-ndjson",
        "application/x-ipynb+json",
    ],
)
def test_safe_file_data_mime_collapses_unrecognised_text(rejected_mime: str) -> None:
    """Text-like MIMEs the Responses API rejects collapse to text/plain.

    The base64 payload doesn't change shape; only the data: URI prefix
    shifts so the provider's MIME validator accepts the block. The
    block's ``filename`` field still carries the original extension
    for the model.
    """
    from omnigent.runtime.content_resolver import _safe_file_data_mime

    assert _safe_file_data_mime(rejected_mime) == "text/plain"


def test_safe_file_data_mime_leaves_non_text_alone() -> None:
    """Non-text MIMEs we don't have an opinion on pass through.

    We don't ship a complete list of non-text formats, so guessing
    "this image/audio/binary type is unsafe" would mislead providers.
    Pass them through and let the provider validate.
    """
    from omnigent.runtime.content_resolver import _safe_file_data_mime

    assert _safe_file_data_mime("image/png") == "image/png"
    assert _safe_file_data_mime("audio/mpeg") == "audio/mpeg"
    assert _safe_file_data_mime("application/zip") == "application/zip"


def test_resolve_yaml_file_uses_text_plain_in_data_uri(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """End-to-end: pasting a YAML file produces ``data:text/plain;...``.

    Regression check for the user-reported "Invalid file data ... got
    unsupported MIME type 'text/yaml'" 400 from the Responses API.
    Without :func:`_safe_file_data_mime` the data URI would carry
    ``text/yaml`` and the provider would reject every turn that
    re-sent the conversation history.
    """
    yaml_content = b"name: example\nvalue: 42\n"
    file_store.files["file_yaml"] = StoredFile(
        id="file_yaml",
        filename="config.yaml",
        bytes=len(yaml_content),
        content_type="application/octet-stream",
        created_at=1000,
    )
    artifact_store.blobs["file_yaml"] = yaml_content

    item = _make_conversation_item(
        [{"type": "input_file", "file_id": "file_yaml", "filename": "config.yaml"}]
    )
    resolved = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )
    resolved_data = resolved[0].data
    assert isinstance(resolved_data, MessageData)
    file_block = resolved_data.content[0]
    assert file_block["file_data"].startswith("data:text/plain;base64,"), (
        f"Expected text/plain data URI for .yaml, got: {file_block['file_data'][:60]}"
    )
    # Filename is still attached so the model sees the original extension.
    assert file_block.get("filename") == "config.yaml"
    # Payload round-trips intact.
    encoded = file_block["file_data"].split(";base64,", 1)[1]
    assert base64.b64decode(encoded) == yaml_content


def test_resolve_image_file_keeps_specific_mime(
    file_store: FakeFileStore,
    artifact_store: FakeArtifactStore,
) -> None:
    """input_image blocks build image_url; the safe-MIME helper must
    not run on that path. The Responses API needs the precise image
    MIME on image_url to pick the right decoder.
    """
    png_bytes = b"\x89PNG\r\n\x1a\n" + b"\x00" * 16  # bare PNG header for the test
    file_store.files["file_png"] = StoredFile(
        id="file_png",
        filename="diagram.png",
        bytes=len(png_bytes),
        content_type="image/png",
        created_at=1000,
    )
    artifact_store.blobs["file_png"] = png_bytes

    item = _make_conversation_item([{"type": "input_image", "file_id": "file_png"}])
    resolved = resolve_content_references(
        [item],
        file_store,
        artifact_store,  # type: ignore[arg-type]
    )
    resolved_data = resolved[0].data
    assert isinstance(resolved_data, MessageData)
    image_block = resolved_data.content[0]
    assert image_block["image_url"].startswith("data:image/png;base64,"), (
        f"Expected image/png data URI, got: {image_block['image_url'][:60]}"
    )


# ── Attachment upload limits ──────────────────────────────────────────


@pytest.mark.parametrize(
    ("content_type", "expected_mb"),
    [
        ("image/png", 5),
        ("image/jpeg", 5),
        ("image/webp", 5),
        ("application/pdf", 20),
        ("text/plain", 10),
        ("text/markdown", 10),
        ("text/x-python", 10),
        ("text/typescript", 10),
        ("application/json", 10),
        ("application/x-ipynb+json", 10),
    ],
)
def test_attachment_upload_limit_allowed_types(content_type: str, expected_mb: int) -> None:
    """Images, PDF, and text-like types get their per-type byte cap."""
    from omnigent.runtime.content_resolver import attachment_upload_limit

    assert attachment_upload_limit(content_type) == expected_mb * 1024 * 1024


@pytest.mark.parametrize(
    "content_type",
    [
        "application/vnd.openxmlformats-officedocument.presentationml.presentation",  # pptx
        "application/vnd.openxmlformats-officedocument.wordprocessingml.document",  # docx
        "application/vnd.ms-excel",
        "application/zip",
        "application/octet-stream",
        "audio/mpeg",
        "video/mp4",
    ],
)
def test_attachment_upload_limit_rejects_unsupported_types(content_type: str) -> None:
    """Office/binary/media types are not uploadable (None ⇒ caller 415s)."""
    from omnigent.runtime.content_resolver import attachment_upload_limit

    assert attachment_upload_limit(content_type) is None


def test_attachment_upload_limits_are_under_global_ceiling() -> None:
    """Every per-type limit stays within the global request-size backstop."""
    from omnigent.runtime.content_resolver import (
        MAX_ATTACHMENT_UPLOAD_BYTES,
        MAX_IMAGE_UPLOAD_BYTES,
        MAX_PDF_UPLOAD_BYTES,
        MAX_TEXT_UPLOAD_BYTES,
    )

    assert MAX_IMAGE_UPLOAD_BYTES <= MAX_ATTACHMENT_UPLOAD_BYTES
    assert MAX_PDF_UPLOAD_BYTES <= MAX_ATTACHMENT_UPLOAD_BYTES
    assert MAX_TEXT_UPLOAD_BYTES <= MAX_ATTACHMENT_UPLOAD_BYTES


@pytest.mark.parametrize(
    ("filename", "expected"),
    [
        ("data.csv", "text/csv"),
        ("notes.txt", "text/plain"),
        ("main.py", "text/x-python"),
        ("app.ts", "text/typescript"),
        ("readme.md", "text/markdown"),
        ("nb.ipynb", "application/x-ipynb+json"),
    ],
)
def test_attachment_text_type_for_extension_recognised(filename: str, expected: str) -> None:
    """Known text/code extensions resolve to a text-like MIME (the fallback
    used when the declared MIME mislabels them as binary)."""
    from omnigent.runtime.content_resolver import attachment_text_type_for_extension

    assert attachment_text_type_for_extension(filename) == expected


@pytest.mark.parametrize(
    "filename",
    ["sheet.xls", "sheet.xlsx", "deck.pptx", "doc.docx", "archive.zip", "blob", None],
)
def test_attachment_text_type_for_extension_rejects_binary(filename: str | None) -> None:
    """Real binaries (and missing/unknown extensions) get no text fallback,
    so they stay rejected even if the declared MIME is wrong."""
    from omnigent.runtime.content_resolver import attachment_text_type_for_extension

    assert attachment_text_type_for_extension(filename) is None


def test_text_code_extensions_resolve_to_allowed_text() -> None:
    """Every declared text/code extension resolves to a text-like type that
    has an upload limit — so the route's extension fallback admits it (no 415),
    regardless of the browser-reported MIME."""
    from omnigent.runtime.content_resolver import (
        _TEXT_CODE_EXTENSIONS,
        attachment_text_type_for_extension,
        attachment_upload_limit,
    )

    for ext in _TEXT_CODE_EXTENSIONS:
        mime = attachment_text_type_for_extension(f"file{ext}")
        assert mime is not None, f"{ext} resolved to no text type"
        assert attachment_upload_limit(mime) is not None, f"{ext} -> {mime} has no limit"


def test_client_server_attachment_extension_parity() -> None:
    """The web client's TEXT_CODE_EXTENSIONS must all be accepted server-side,
    even when the browser reports a non-text MIME — the parity contract the two
    share. Guards against the client gate admitting a file the upload route then
    415s (the divergence Polly flagged)."""
    import re
    from pathlib import Path

    from omnigent.runtime.content_resolver import (
        _resolve_content_type,
        attachment_text_type_for_extension,
        attachment_upload_limit,
    )

    ts_path = Path(__file__).resolve().parents[2] / "web" / "src" / "lib" / "attachments.ts"
    if not ts_path.exists():
        pytest.skip("web/src/lib/attachments.ts not present (server-only checkout)")
    block = ts_path.read_text().split("TEXT_CODE_EXTENSIONS = new Set([")[1].split("]")[0]
    client_exts = re.findall(r'"(\.[a-z0-9]+)"', block)
    assert client_exts, "could not parse client TEXT_CODE_EXTENSIONS"

    # MIMEs a browser/OS might attach to these extensions, including wrong ones.
    worst_case_mimes = [
        "",
        "application/octet-stream",
        "video/mp2t",  # .ts
        "application/xml",  # .xml
        "application/x-ruby",  # .rb
    ]

    def server_accepts(name: str, browser_mime: str) -> bool:
        content_type = _resolve_content_type(browser_mime, name)
        limit = attachment_upload_limit(content_type)
        if limit is None:
            ext_type = attachment_text_type_for_extension(name)
            if ext_type is not None:
                limit = attachment_upload_limit(ext_type)
        return limit is not None

    rejected = [
        (ext, mime)
        for ext in client_exts
        for mime in worst_case_mimes
        if not server_accepts(f"file{ext}", mime)
    ]
    assert not rejected, f"client accepts but server would 415: {rejected}"