fix(backends): reject empty user content before sending chat request

mellea/backends/litellm.py

@@ -42,6 +42,7 @@
     chat_completion_delta_merge,
     extract_model_tool_requests,
     get_current_event_loop,
+    has_user_content,
     message_to_openai_message,
     prefetch_audio_urls,
     send_to_queue,
@@ -370,6 +371,17 @@ async def _generate_from_chat_context_standard(
             case _:
                 messages.extend(self.formatter.to_chat_messages([action]))
 
+        # Issue #1597: refuse to send an empty user prompt; see
+        # `has_user_content` for why whitespace-only text still counts as empty.
+        if not has_user_content(messages):
+            raise ValueError(
+                "Refusing to call the model: no user-role content in the assembled "
+                "conversation. This usually means a stateless context (e.g. "
+                "SimpleContext) was combined with an empty or whitespace-only "
+                "action; recorded turns are not forwarded to the model. See "
+                "issue #1597."
+            )
+
         # TODO: the supports_vision function is not reliably predicting if models support vision. E.g., ollama/llava is not a vision model?
         # if any(m.images is not None for m in messages):
         #     # check if model can handle images

mellea/backends/ollama.py

@@ -37,6 +37,7 @@
     DEFAULT_CHUNK_TIMEOUT,
     ClientCache,
     get_current_event_loop,
+    has_user_content,
     merge_provider_fields,
     message_to_openai_message,
     messages_to_docs,
@@ -1076,6 +1077,8 @@ async def generate_from_chat_context(
                 cannot be downloaded or decoded.
             ValueError: If a message contains an `AudioBlock` or `AudioUrlBlock`;
                 Ollama does not support audio input.
+            ValueError: If no user-role message has non-whitespace text,
+                images, audio, or documents.
         """
         # Start by awaiting any necessary computation.
         await self.do_generate_walk(action)
@@ -1090,6 +1093,17 @@ async def generate_from_chat_context(
         messages: list[Message] = self.formatter.to_chat_messages(linearized_context)
         # Add the final message.
         messages.extend(self.formatter.to_chat_messages([action]))
+
+        # Issue #1597: refuse to send an empty user prompt; see
+        # `has_user_content` for why whitespace-only text still counts as empty.
+        if not has_user_content(messages):
+            raise ValueError(
+                "Refusing to call the model: no user-role content in the assembled "
+                "conversation. This usually means a stateless context (e.g. "
+                "SimpleContext) was combined with an empty or whitespace-only "
+                "action; recorded turns are not forwarded to the model. See "
+                "issue #1597."
+            )
         # construct the conversation from our messages, adding a system prompt at the first message if one was provided.
         conversation: list[dict] = []
         # We use system prompt None/empty-string semantics in a way that is consistent with Hugging Face and other libraries.

mellea/backends/openai.py

@@ -45,6 +45,7 @@
     chat_completion_delta_merge,
     extract_model_tool_requests,
     get_current_event_loop,
+    has_user_content,
     is_vllm_server_with_structured_output,
     message_to_openai_message,
     messages_to_docs,
@@ -1366,6 +1367,10 @@ async def generate_from_chat_context(
         Returns:
             tuple[ModelOutputThunk[C], Context]: A thunk holding the (lazy) model output
                 and an updated context that includes `action` and the new output.
+
+        Raises:
+            ValueError: If no user-role message has non-whitespace text,
+                images, audio, or documents.
         """
         await self.do_generate_walk(action)
 
@@ -1401,6 +1406,17 @@ async def _generate_from_chat_context_standard(
         # ALoraRequirement may arrive here when no adapter is registered;
         # _generate is responsible for logging a warning in that case.
 
+        # Issue #1597: refuse to send an empty user prompt; see
+        # `has_user_content` for why whitespace-only text still counts as empty.
+        if not has_user_content(messages):
+            raise ValueError(
+                "Refusing to call the model: no user-role content in the assembled "
+                "conversation. This usually means a stateless context (e.g. "
+                "SimpleContext) was combined with an empty or whitespace-only "
+                "action; recorded turns are not forwarded to the model. See "
+                "issue #1597."
+            )
+
         conversation: list[dict] = []
 
         # Resolve any audio URLs off-thread so the sync serializer below hits the cache

mellea/helpers/__init__.py

@@ -24,6 +24,7 @@
     OPENAI_COMPATIBLE_WIRE_PROVIDERS,
     chat_completion_delta_merge,
     extract_model_tool_requests,
+    has_user_content,
     merge_provider_fields,
     message_to_openai_message,
     messages_to_docs,

mellea/helpers/openai_compatible_helpers.py

@@ -618,3 +618,32 @@ def build_tool_calls(output: ModelOutputThunk) -> list[ToolCallDict] | None:
         tool_calls.append(tool_call)
 
     return tool_calls
+
+
+def has_user_content(messages: list[Message]) -> bool:
+    """Whether the assembled conversation carries real user-role content.
+
+    Issue #1597: `SimpleContext` intentionally discards recorded turns from
+    `view_for_generation()`, so a caller who chains `.add(...)` and then
+    passes an empty action would otherwise hit the model with no user-role
+    content at all. Some chat models (e.g. Granite 4.2, see #1587) spin on
+    empty prompts and burn tokens silently. Whitespace-only text counts as
+    empty; images, audio, and documents count as content.
+
+    Args:
+        messages: The chat messages about to be sent to the model.
+
+    Returns:
+        True if any user-role message has non-whitespace text, images,
+        audio, or documents; False otherwise.
+    """
+    return any(
+        m.role == "user"
+        and (
+            (m.content and m.content.strip())
+            or m.images
+            or m.audio
+            or getattr(m, "_docs", None)
+        )
+        for m in messages
+    )

mellea/stdlib/context/simple.py

@@ -11,11 +11,29 @@
 
 
 class SimpleContext(Context):
-    """A `SimpleContext` is a context in which each interaction is a separate and independent turn. The history of all previous turns is NOT saved.."""
+    """A `SimpleContext` is a context in which each interaction is a separate and independent turn. The history of all previous turns is NOT saved..
+
+    Note:
+        Because `view_for_generation` always returns an empty list, anything
+        passed to `SimpleContext.add` is **never forwarded to the model** — it
+        is recorded only on the in-memory context chain. The action passed to
+        `generate_from_context` (or `MelleaSession.chat`) is the *only* thing
+        that reaches the model. Combining `.add(...)` with an empty/whitespace
+        action therefore produces an empty user prompt; the OpenAI, LiteLLM,
+        and Ollama backends now reject such calls with a `ValueError` (see
+        issue #1597) rather than sending an empty conversation to the model.
+    """
 
     def add(self, c: Span) -> Self:
         """Add a new component or CBlock to the context and return the updated context.
 
+        The added span is stored on the context chain but is **not forwarded
+        to the model on subsequent generations** — `SimpleContext.view_for_generation`
+        always returns an empty list, so each generation is treated as a
+        stateless, independent turn. To actually talk to the model, pass the
+        prompt as the `action` argument to `MelleaSession.chat` /
+        `Backend.generate_from_context`, not via `add`.
+
         Args:
             c (Span): The component, content
                 block, or model output to record.
@@ -33,7 +51,9 @@ def view_for_generation(self) -> list[Span] | None:
         """Return an empty list, since `SimpleContext` does not pass history to the model.
 
         Each call to the model is treated as a stateless, independent exchange.
-        No prior turns are forwarded.
+        No prior turns are forwarded. Spans recorded via `add` are kept on
+        the in-memory chain (`as_list`) for inspection but discarded for
+        generation.
 
         Returns:
             list[Span] | None: Always an empty list.