LocalHFBackend silently answers a blank prompt when SimpleContext is given an empty action

[planetf1]

What goes wrong

#1632 makes the OpenAI, LiteLLM, and Ollama backends raise ValueError when the assembled conversation has no user content. That happens when a caller records turns on a SimpleContext and then generates with an empty action (#1597). LocalHFBackend has no such check. The same call sends the model one empty user turn, raises nothing, and returns an answer to a prompt the caller never wrote.

With ibm-granite/granite-4.2-3b the recorded turn is dropped and the model reasons about a message that doesn't exist:

import asyncio

from mellea.backends.huggingface import LocalHFBackend
from mellea.core import CBlock
from mellea.stdlib.context import SimpleContext


async def main():
    backend = LocalHFBackend(model_id="ibm-granite/granite-4.2-3b")
    ctx = SimpleContext().add(CBlock("Summarise this report: revenue rose 12% in Q3."))
    mot, _ = await backend.generate_from_context(CBlock(""), ctx)
    print(repr(await mot.avalue()))


asyncio.run(main())

Output:

'Okay, the user sent a message that\'s just the word "hello". I need to respond appropriately'

Expected: the same ValueError the other backends raise.

Mechanism

Both HF chat paths, standard (mellea/backends/huggingface.py:1782) and KV-cache (:1587), build messages through to_chat() in mellea/backends/utils.py:74. It takes ctx.view_for_generation(), which is empty for SimpleContext by design, appends the action, and hands the result to apply_chat_template without checking for user content. SimpleContext inherits _is_chat_context = True (mellea/core/base.py:2066), so it takes this path.

Capturing the conversation passed to apply_chat_template shows [{'role': 'user', 'content': ''}] on both HuggingFaceTB/SmolLM2-135M-Instruct and ibm-granite/granite-4.2-3b. A non-empty prompt through the same capture shows the expected user turn.

Suggested fix

Call has_user_content() in to_chat() straight after the action is appended and raise the same ValueError as the other backends. One call covers both HF paths. HF already rejects images and audio via _check_no_multimodal_blocks, so only the text and documents branches apply.

has_user_content() lives in mellea/helpers/openai_compatible_helpers.py. Once a non-OpenAI-compatible backend calls it, a backend-neutral helpers module is a better home.

Test: mirror test/backends/test_simple_context_guard.py, with SimpleContext + CBlock("") raising before generation, plus a documents pass-through case.

Scope and limits

• Out of scope: WatsonxAIBackend (deprecated) and _generate_from_intrinsic, which doesn't append the action.
• The token-burning stall from feat: add Granite 4.2 model defaults #1587 didn't reproduce here. Granite 4.2-3b answered the blank prompt in 74 tokens against a 1024-token cap (seed 0, one run). On HF the cost is a silently wrong answer.