Fix/prune empty assistant messages

[mmastrangelo]

Summary

Prevent ActionStep.to_messages() from emitting empty assistant messages when model_output is empty or whitespace-only.

Motivation / Problem

In tool-calling flows, models may return tool calls without any assistant text (tool-call-only responses). In these cases ActionStep.model_output can be "" or whitespace. The previous implementation always appended an assistant ChatMessage built from model_output.strip(), which could result in an assistant message with empty text content.

This causes:

• Noise / unnecessary tokens in the replayed conversation history.
• Provider compatibility issues: in a live agent setup using xAI Grok (grok-code-fast-1), consecutive empty assistant messages were correlated with a reproducible provider-side token counting failure (“maximum prompt length” / bogus token count), despite otherwise reasonable payload sizes.

Reproduction

1. Run a tool-calling agent against a model/provider that sometimes emits tool calls with no assistant text.
2. On those steps, ActionStep.model_output may be "" or whitespace.
3. When converting memory back to chat messages, ActionStep.to_messages() appended an assistant message using model_output.strip().
4. If strip() produced an empty string, an empty assistant message was added to the conversation history.
5. In the reported grok-code-fast-1 setup, these empty assistant messages were sufficient to reliably trigger a provider token-counting failure.

Change

In src/smolagents/memory.py, ActionStep.to_messages() now:

• computes text = self.model_output.strip()
• only appends the assistant ChatMessage when text is non-empty

This preserves behavior for normal non-empty outputs while preventing informationless assistant messages.

Tests

• Added unit test: tests/test_memory.py::test_action_step_to_messages_ignores_blank_model_output
  • Ensures blank/whitespace-only model_output does not produce an assistant message.

Validation / Test Plan

• Run unit tests (make test) and quality checks (make quality).
• Confirmed locally with a live agent setup using grok-code-1 that the previously reproducible token computation error no longer occurs after this change.