Files
LLeMbas/tests/test_images_chat.py
T
Jaroslav Beneš 47d1ddbc3c Draw a picture, on a ComfyUI you are running
The last unbuilt capability, and built the way CLAUDE.md said it had to be: a
ToolDef reaching resolve_tools plus a permission and a capability flag, not a new
code path. The only genuinely new UI is one branch in the transcript.

services/images/ is three modules. comfy.py speaks HTTP -- submit, poll /history,
fetch the PNG, /free, and an /object_info discovery for the admin page only.
Polled and not socketed, because holding a connection open for the length of a
generation is the live-connection state the whole ssh.py design forbids, and the
thing being waited for takes tens of seconds anyway. The base URL is exempt from
the SSRF guard by construction, exactly as Connection.base_url and the audio
endpoints are -- said out loud in the docstring, because a default of
127.0.0.1:8188 is precisely the shape that guard exists to refuse and therefore
reads as a hole rather than a decision.

workflow.py fills a template, and the one thing that matters is that it walks the
parsed JSON rather than the text of it. A value that is exactly "{{steps}}"
becomes the number 20; ComfyUI validates types and refuses the string. A
placeholder inside a longer string is still text, which is what makes
"{{prompt}}, masterpiece" work -- and text substitution would additionally mean a
prompt containing a quotation mark produced a document that no longer parses, on
the one input guaranteed to hold arbitrary text. Which node holds the prompt is
the administrator's statement rather than a guess from node types: sniffing for
the first CLIPTextEncode works on the shipped workflow and on nothing else, and
swaps positive for negative the first time somebody reorders them. seed has no
fixed default, because one would make every unspecified generation identical and
make the retry loop redraw the same rejected picture four times.

tool.py is one call, one finished image. Returning every attempt to the
conversation would cost a round each, make the ceiling advisory rather than
enforced, and walk the reader past every reject -- so the reviewer lives inside
the tool and is asked about *bytes*: an attempt about to be discarded should not
leave an Attachment behind, so it sees a downscaled preview built in memory and
only the kept image is written. Anything that goes wrong in review is a keep;
losing a picture because a judging request timed out would be the check
destroying the thing it was checking. The last attempt is kept whatever the
verdict, so a request always produces something. Rejects are recorded, not
stored.

Preserve VRAM unloads the chat's own connection and nothing else, because the
memory being freed belongs to one machine: local llama-swap answers GET /unload,
and a box on the network has no reason to be unloaded when ComfyUI wants memory
here. The swap goes round the review rather than round the tool, which costs two
model loads per retry -- so the two settings are independent and the page warns
when both are on. Nothing loads the LLM back: the reply's next request does, and
that step exists in the description and not in the code, so the code says so.

Two rules elsewhere had to be drawn for the first time. message_payload sends
images only on user turns -- no assistant message had ever carried one, and the
moment one does the multimodal list form on an assistant turn is rejected by
OpenAI and most local runners, breaking every later turn in the chat. And
files.store gained keep_original, because _process_image turns anything without
alpha into JPEG q85 at 1400px: right for a phone photo, a visible loss on the one
output this feature exists to produce.

/image sends the ordinary message with force_tool, which becomes tool_choice for
the first round only -- left in place the reply would draw a picture, be asked
again, and draw another. FORCEABLE_TOOLS is an allow list because the name is
read off a form.

ToolContext gained chat_id, and that fixed a tool nobody had ever successfully
run: _run_scratch_write read context.chat_id on a dataclass with no such field,
so every call raised AttributeError, swallowed by run_tool's blanket except into
"the scratch_write tool failed" -- indistinguishable from a model calling it
wrongly. The test that existed asserted the family and the risk, which are
properties of the declaration rather than of the code.

Verified against the real ComfyUI 0.27.0 on this machine rather than against
documentation: every endpoint shape here was read off it, a generation ran end to
end through the client, the reviewer was shown a matching and a mismatched prompt
and answered KEEP and RETRY correctly, and the unload hook fired for the local
llama-swap and not for the remote box.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-05 14:13:19 +02:00

298 lines
11 KiB
Python

"""Image generation where it meets the conversation.
Three seams, each of which fails silently if it is got wrong: how a generated
image is bound to the reply that made it, why it is never replayed on an
assistant turn, and what `/image` actually sends.
"""
from __future__ import annotations
from pathlib import Path
import pytest
from fastapi.testclient import TestClient
from sqlalchemy import select
import lembas
from lembas.db.models import ROLE_ASSISTANT, ROLE_USER, Attachment, Chat, Connection, Message, Model
from lembas.services import chat as chat_service
from lembas.services import files as files_service
from lembas.services import generation as generation_service
from lembas.services.crypto import encrypt
PNG = (
b"\x89PNG\r\n\x1a\n\x00\x00\x00\rIHDR\x00\x00\x00\x01\x00\x00\x00\x01\x08\x02\x00\x00\x00"
b"\x90wS\xde\x00\x00\x00\x0cIDATx\x9cc```\x00\x00\x00\x04\x00\x01\xf6\x178U\x00\x00\x00"
b"\x00IEND\xaeB`\x82"
)
@pytest.fixture
def vision_chat(db, user_id):
connection = Connection(name="c", base_url="http://127.0.0.1:1", api_key_encrypted=encrypt(""))
db.add(connection)
db.commit()
db.add(
Model(
connection_id=connection.id,
model_id="m",
capabilities_json={"vision": True, "tools": True},
)
)
chat = Chat(user_id=user_id, model_id="m", connection_id=connection.id)
db.add(chat)
db.commit()
return chat
# --- Storing what was drawn ----------------------------------------------------
def test_a_generated_image_is_kept_as_it_arrived(db, user_id, vision_chat):
"""`_process_image` transcodes to JPEG q85 and downscales to 1400px, which
is right for a phone photo and a visible loss on the one output this feature
exists to produce."""
attachment = files_service.store(
db,
user_id=user_id,
chat_id=vision_chat.id,
payload=PNG,
filename="out.png",
keep_original=True,
)
assert attachment.media_type == "image/png"
assert files_service.stored_path(attachment.stored_name).read_bytes() == PNG
def test_an_ordinary_upload_is_still_processed(db, user_id, vision_chat):
"""The flag is opt-in, and the protection it skips still applies to
everything that arrives from outside."""
attachment = files_service.store(
db, user_id=user_id, chat_id=vision_chat.id, payload=PNG, filename="photo.png"
)
assert attachment.media_type == "image/jpeg"
def test_a_corrupt_image_is_still_refused(db, user_id, vision_chat):
"""What `keep_original` skips is the resize and the transcode, not the
check that this is an image at all."""
with pytest.raises(files_service.FileError):
files_service.store(
db,
user_id=user_id,
chat_id=vision_chat.id,
payload=b"\x89PNG\r\n\x1a\n" + b"rubbish",
filename="broken.png",
keep_original=True,
)
def test_the_loop_binds_the_image_to_the_reply(db, user_id, vision_chat):
"""A runner cannot write the message row -- `_persist` is the single writer
-- so the runner makes the attachment and the loop says which turn owns it,
exactly as it already does for a canvas tab."""
message = chat_service.create_message(db, vision_chat, ROLE_ASSISTANT, "here it is")
attachment = files_service.store(
db, user_id=user_id, chat_id=vision_chat.id, payload=PNG, filename="a.png"
)
generation_service._bind_attachments(db, vision_chat, message, [attachment.id])
db.commit()
db.refresh(attachment)
assert attachment.message_id == message.id
def test_binding_refuses_a_row_from_another_chat(db, user_id, vision_chat):
"""The ids arrive on a tool event, which is a dict a runner built. Without
the narrowing a forged one would pull somebody else's file into this
conversation -- the reason `files.claim` is scoped the same way."""
other = Chat(user_id=user_id, model_id="m")
db.add(other)
db.commit()
elsewhere = files_service.store(
db, user_id=user_id, chat_id=other.id, payload=PNG, filename="a.png"
)
message = chat_service.create_message(db, vision_chat, ROLE_ASSISTANT, "x")
generation_service._bind_attachments(db, vision_chat, message, [elsewhere.id])
db.commit()
db.refresh(elsewhere)
assert elsewhere.message_id is None
def test_binding_refuses_a_row_already_bound(db, user_id, vision_chat):
first = chat_service.create_message(db, vision_chat, ROLE_ASSISTANT, "one")
second = chat_service.create_message(db, vision_chat, ROLE_ASSISTANT, "two")
attachment = files_service.store(
db,
user_id=user_id,
chat_id=vision_chat.id,
payload=PNG,
filename="a.png",
message_id=first.id,
)
generation_service._bind_attachments(db, vision_chat, second, [attachment.id])
db.commit()
db.refresh(attachment)
assert attachment.message_id == first.id
# --- What reaches the model ----------------------------------------------------
def test_an_image_on_an_assistant_turn_is_never_replayed(db, user_id, vision_chat):
"""The multimodal list form on an `assistant` turn is rejected outright by
OpenAI and by most local runners -- and it would break not that turn but
every later one in the chat."""
reply = chat_service.create_message(db, vision_chat, ROLE_ASSISTANT, "here it is")
files_service.store(
db,
user_id=user_id,
chat_id=vision_chat.id,
payload=PNG,
filename="a.png",
message_id=reply.id,
)
db.refresh(reply)
payload = chat_service.message_payload(reply, vision=True)
assert isinstance(payload["content"], str), "no image_url parts on an assistant turn"
assert payload["content"] == "here it is"
def test_an_image_a_person_sent_still_reaches_the_model(db, user_id, vision_chat):
"""The rule narrows assistant turns and nothing else."""
turn = chat_service.create_message(db, vision_chat, ROLE_USER, "what is this?")
files_service.store(
db,
user_id=user_id,
chat_id=vision_chat.id,
payload=PNG,
filename="a.png",
message_id=turn.id,
)
db.refresh(turn)
payload = chat_service.message_payload(turn, vision=True)
assert isinstance(payload["content"], list)
assert any(part["type"] == "image_url" for part in payload["content"])
def test_a_generated_image_still_renders_in_the_bubble(
client: TestClient, db, user_id, vision_chat
):
"""It is an attachment on the assistant message, and `_message.html` renders
attachments for either role -- so the reader sees it without the template
learning anything new."""
reply = chat_service.create_message(db, vision_chat, ROLE_ASSISTANT, "here it is")
attachment = files_service.store(
db,
user_id=user_id,
chat_id=vision_chat.id,
payload=PNG,
filename="a.png",
message_id=reply.id,
)
page = client.get(f"/chat/{vision_chat.id}").text
assert f"/api/files/{attachment.id}/content" in page
# --- Forcing the tool ----------------------------------------------------------
def test_the_forced_tool_reaches_the_request(db, vision_chat):
tools = [{"type": "function", "function": {"name": "image_generate", "parameters": {}}}]
body = chat_service.build_request(db, vision_chat, tools=tools, force_tool="image_generate")
assert body["tool_choice"] == {
"type": "function",
"function": {"name": "image_generate"},
}
def test_nothing_is_forced_by_default(db, vision_chat):
"""A provider strict about unknown parameters must see exactly the request
it always did until somebody types a slash command."""
tools = [{"type": "function", "function": {"name": "image_generate", "parameters": {}}}]
assert "tool_choice" not in chat_service.build_request(db, vision_chat, tools=tools)
def test_a_tool_that_was_not_offered_cannot_be_forced(db, vision_chat):
"""`resolve_tools` still decides what exists. Forcing something absent from
the array is a request most endpoints reject outright."""
tools = [{"type": "function", "function": {"name": "web_search", "parameters": {}}}]
body = chat_service.build_request(db, vision_chat, tools=tools, force_tool="image_generate")
assert "tool_choice" not in body
def test_forcing_needs_a_tools_array_at_all(db, vision_chat):
assert "tool_choice" not in chat_service.build_request(
db, vision_chat, force_tool="image_generate"
)
def test_the_endpoint_only_accepts_names_from_the_allow_list(
client: TestClient, db, registered, vision_chat, monkeypatch
):
"""This becomes `tool_choice`, so a name read straight off a form would let
anyone who can send a message decide what the model must do next."""
from lembas.api import chats as chats_api
seen: dict = {}
monkeypatch.setattr(
chats_api.generation_service,
"ensure",
lambda chat_id, message_id, *, force_tool="": seen.update(force_tool=force_tool),
)
client.post(
f"/api/chats/{vision_chat.id}/messages",
data={"content": "hello", "force_tool": "shell_run"},
)
assert seen["force_tool"] == "", "not on the list, so not forced"
# That first turn left an unfinished assistant row behind, and a second
# message while one is in flight is *queued* rather than sent -- so it would
# never reach `ensure` at all. Finish it first.
db.query(Message).filter(Message.complete.is_(False)).update({"complete": True})
db.commit()
client.post(
f"/api/chats/{vision_chat.id}/messages",
data={"content": "a bicycle", "force_tool": "image_generate"},
)
assert seen["force_tool"] == "image_generate"
def test_the_image_command_is_in_the_table():
"""`/help` reads this list, so a command missing from it is one nobody can
discover -- the direction this actually rots."""
source = (Path(lembas.__file__).parent / "web/static/js/commands.js").read_text(
encoding="utf-8"
)
assert 'name: "image"' in source
assert 'body.append("force_tool", "image_generate")' in source
assert "htmx.process" in source, "or the reply's sse-connect is inert markup"
def test_a_forced_round_does_not_force_the_next_one(db, vision_chat):
"""Leaving `tool_choice` in place would make every round call the tool
again: draw a picture, be asked again, draw another."""
source = (Path(lembas.__file__).parent / "services/generation.py").read_text(encoding="utf-8")
assert 'payload.pop("tool_choice", None)' in source
def test_a_queued_turn_is_not_lost_when_it_was_forced(db, user_id, vision_chat):
"""`/image` typed while a reply is streaming queues like anything else. The
forcing is on the generation, so a queued turn simply arrives unforced --
which is right: by then the model has the words and the context."""
db.add(Message(chat_id=vision_chat.id, role=ROLE_ASSISTANT, content="", complete=False))
db.commit()
from lembas.api import chats as chats_api
assert chats_api._reply_in_flight(db, vision_chat) is True
assert db.scalar(select(Attachment)) is None